Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-NVFP4-NInfer

A model post-trained with several rounds of SFT and RLOO on top of Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2, which is itself the official Qwen3.8-27B post-trained with SFT and SimPO. It is built to fix the original model's habit of failing to stop: it already has an answer, yet keeps re-deriving until it hits the 94K cap with no final answer. This repository provides the three NVFP4 tiers, NVFP4 W4A16 (NVFP4/W4A16/), NVFP4 W4A4 (NVFP4/W4A4/), and NVFP4 W4A4-W8A8 (NVFP4/W4A4-W8A8/, mixed precision)—all fully multimodal (text + image/video) with the official BF16 MTP bundled and their own DFlash2 draft (DFlash2-FP8/) inside each tier directory, loading in SGLang without patches—plus two official NInfer tiers, NInfer W4A4 (NVFP4-NInfer/W4A4/) and NInfer W4A4-W8A8 (NVFP4-NInfer/W4A4-W8A8/), each a single .ninfer package with a Q8 MTP head and the DFlash2 draft built in, loadable only by the NInfer engine, INT8 W8A8 (INT8/W8A8/), text + image, loadable only by vLLM, and INT4 W4A16 (INT4/W4A16/), text + image with a built-in BF16 MTP head, loadable only by vLLM. The BF16 and static FP8 tiers are in the main repository Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2.

• Far fewer 94K truncations: under the same protocol, 4 → 1 on GPQA and 13 → 3 on LCB.
• No regression on any suite: GPQA 178/198, MMLU 450/500, LCB 90/100 (static FP8), vs. 177 / 444 / 83 for the original Qwen3.8 27B (FP8).
• Lossless quantization: GPQA and LCB are identical across BF16, dynamic FP8, and static FP8.
• Two speculative-decoding options (mutually exclusive): about 180 tok/s per request with DFlash2 and 91 tok/s with MTP, vs. about 46 tok/s without speculation.

Related repositories

• Main repository (BF16 / FP8 / NVFP4 / INT8 / GGUF and all other tiers): Hugging Face · ModelScope
• GGUF / GGUF-NInfer (llama.cpp, NInfer, built-in MTP): Hugging Face · ModelScope

Training method

Lineage: original Qwen3.8 27B (FP8; 177 / 444 / 83) → our SFT + SimPO110 final, i.e. the EfficientThink base (171 / 442 / 89) → K3 continuation SFT → week-2 SFT (week2dose, update-225) → first RLOO round (182 groups) → another SFT round (merge-sft-100), giving sft-base-rloo (177 / 448 / 90) → second RLOO round (172 groups) → Coder390 (178 / 445 / 90, dynamic FP8). Numbers are GPQA / MMLU / LCB, all full suites under the same 100K protocol.

Training base: this model starts from our own EfficientThink SFT + SimPO model (the SimPO110 final) and goes through several rounds of SFT and RLOO, which is itself a post-trained version of the official Qwen3.8-27B: Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2.

The problem: after the model already has a local answer, it keeps re-running the same derivation with Wait / Actually until the context is full, and the final answer is empty. In most of these cases the model can solve the problem; it just does not stop.

RLOO data: 8 trajectories sampled per problem; only groups with both correct and wrong trajectories are kept (all-correct groups are dropped). Groups with 0 or 1 correct trajectory receive one reviewed short teacher trajectory, which replaces the shortest wrong trajectory in that group (27 groups). Final set: 172 groups, 1,376 trajectories: 859 correct and 517 wrong, including 68 empty answers.

Reward: penalties apply only to wrong trajectories; long correct reasoning still earns a positive reward, so necessary long reasoning is not suppressed.

Result: under the same protocol, 94K truncations drop from 4 to 1 on GPQA and from 13 to 3 on LCB (FP8) compared with the original Qwen3.8 27B (FP8), and none of the three scores falls below the original (all FP8 figures; see the NVFP4 scores below).

Scores

Protocol: NVFP4 quantizations of the same Coder390 merged weights; one RTX PRO 6000, C8, 100K context, 94,208-token generation cap, no client-side timeout, DFlash2 draft, SGLang; full GPQA 198, MMLU 500, and LCB 100. Empty answers count as wrong, truncated-but-correct answers count as correct, and every failed sample stays in the denominator.

LCB correctness is judged by each problem's pass field. Raw records: evaluation/SCORES_NVFP4_W4A16.json, evaluation/SCORES_NVFP4_W4A4.json, and evaluation/SCORES_NVFP4_MIXED.json (the W4A4-W8A8 tier). The Coder390 static FP8 column is taken from the main repository (two RTX PRO 6000 GPUs, C8 per GPU; raw record evaluation/SCORES_STATICFP8.json in the main repository), for comparison only.

NVFP4 tiers: scores, reasoning length, 94K truncation, and empty answers

Reasoning length is usage.reasoning_tokens; a 94K truncation means the output hit the generation cap. LCB empty answers are the number of problems for which no runnable code could be extracted. The original Qwen3.8 27B (FP8) and Coder390 static FP8 rows use two RTX PRO 6000 GPUs with C8 per GPU (taken from the main repository, for comparison only); the three NVFP4 tiers use a single RTX PRO 6000 at C8; the rest of the protocol is the same.

Bold marks values better than the original Qwen3.8 27B (FP8): a higher score, or lower P50 / P70 / P90, 94K truncations, or empty answers; values equal to or worse than the original are not bolded.

LCB improves the most: all 13 94K truncations of the original Qwen3.8 27B (FP8) ended with empty code (13 truncations, 13 empty answers); Coder390 static FP8 has 3, and the three NVFP4 tiers cut this to 2–4. The 3 for W4A4 are problems 0, 53, and 96 (problem ids in the score file), likewise truncated with no code; of the 4 for W4A16, problems 6, 7, and 54 have no code, and problem 53 has 20 characters of code but is wrong; of the 2 for W4A4-W8A8, problem 11 has no code, and problem 6 has 2,338 characters of code but is wrong. On GPQA the original had 4 truncations at the 94K cap, Coder390 static FP8 has 1, and each NVFP4 tier has 2; on MMLU the original had 1, Coder390 static FP8 has 1, and the three tiers have 0–2, with no clear change.

Buckets are "correct / questions in bucket", split by reasoning length, left-closed and right-open.

NInfer tiers: scores, reasoning length, 94K truncations, and empty answers

Protocol: scores carried over from the official triad run on 2026-10-05 on the same NInfer text path; one RTX PRO 6000, official NInfer engine (Neroued/ninfer), C8, 100K context, 94,208 generation cap, client not timed, no speculation; full GPQA 198 / MMLU 500 / LCB 100; empty answers count as wrong, truncations that still answer correctly count as correct, and every failure stays in the denominator. The packages uploaded now add a Q8 MTP head, the DFlash2 draft, and the proposal head on top of that text path; the text path is unchanged and the full triad was not re-run. Only the NInfer engine can load .ninfer; SGLang / vLLM / llama.cpp / transformers cannot.

Reasoning length is usage.reasoning_tokens (i.e. completion_tokens_details.reasoning_tokens); a 94K truncation means the generation cap was hit. LCB empty answers are the number of problems for which no runnable code could be extracted. Bold means better than the original Qwen3.8 27B (FP8).

For NInfer W4A4, the 2 LCB empty answers are problems 11 and 15, both truncated with no code; the 2 GPQA empty answers are problems 79 and 127, both truncated with an empty pred. MMLU has no truncations and no empty answers. For NInfer W4A4-W8A8, the 3 LCB empty answers are problems 17, 53, and 54 (problem 17 stopped normally with no code; 53 and 54 were truncated with no code); the 2 GPQA empty answers are problems 79 and 127, both truncated with an empty pred. MMLU has no truncations and no empty answers.

Buckets are "correct / problems in bucket", half-open on the left.

INT8 W8A8: scores, 94K truncations, and empty answers

Protocol: one RTX PRO 6000, vLLM 0.28, C8, 100K context, 94,208 generation cap, client not timed, no speculation, reasoning effort xhigh; full GPQA 198 / MMLU 500 / LCB 100; empty answers count as wrong, truncations that still answer correctly count as correct, and every failure stays in the denominator. INT8 W8A8 loads only in vLLM.

A 94K truncation means the generation cap was hit. LCB empty answers are the number of problems for which no runnable code could be extracted. Bold means better than the original Qwen3.8 27B (FP8); its numbers are in the table above.

Note: the INT8 W8A8 eval ran vLLM without a reasoning parser, so usage.reasoning_tokens is 0 throughout and reasoning plus answer were returned as content. GPQA and MMLU reasoning lengths were recounted from the raw text: the text before </think> in the content, counted with the BF16 tokenizer (the same method as the GGUF tiers). The LCB result file kept only the code, not the reasoning text, so it cannot be counted; that cell reads "—".

GPQA's 1 truncation is problem 127, with an empty pred, so it also counts as the empty answer. LCB's 2 truncations are problems 7 and 53: problem 7 left code that raised a runtime error and was judged wrong; problem 53 was truncated, but its code was still judged correct; LCB has no empty answers. MMLU has no truncations and no empty answers. By difficulty, LCB is hard 38/46, medium 30/31, and easy 23/23.

INT4 W4A16: scores, 94K truncations, and empty answers

Protocol: one RTX PRO 6000, vLLM 0.28, C8, 100K context, 94,208 generation cap, client not timed, reasoning effort xhigh, MTP speculation with 3 draft tokens (built-in head), --reasoning-parser qwen3; full GPQA 198 / MMLU 500 / LCB 100; empty answers count as wrong, truncations that still answer correctly count as correct, and every failure stays in the denominator. The run was paused on 2026-10-09 and resumed with the same server settings; finished problems were kept.

A 94K truncation means the generation cap was hit. LCB empty answers are the number of problems for which no runnable code could be extracted. Bold means better than the original Qwen3.8 27B (FP8); its numbers are in the all-tier table above.

Reasoning lengths come from usage.reasoning_tokens returned by vLLM. GPQA: 1 truncation (problem 79), 1 empty answer (problem 79). MMLU has no truncations and no empty answers. LCB: 2 truncations (problem atcoder:arc182_e:6, atcoder:arc195_d:54), 2 empty answers (problem atcoder:arc182_e:6, atcoder:arc195_d:54). By difficulty, LCB is hard 37/46, medium 27/31, and easy 23/23. Mean MTP accept length over the run (vLLM log, 3 draft tokens): 2.81.

Quantization tiers

BPW (bits per weight) is the effective average bit width of the whole package: the total bytes of the main-model weight files (.safetensors, including the vision tower, MTP, and scale tensors, excluding the DFlash2 draft) × 8 ÷ total parameter count. The parameter count is taken from the safetensors headers and is identical across the three tiers at 27,781,427,952 (an NVFP4 packed U8 tensor counts 2 elements per byte; scale tensors add no elements). The weight-file totals for NVFP4 W4A16, NVFP4 W4A4, and NVFP4 W4A4-W8A8 are 20,593,097,872, 20,593,145,056, and 23,749,329,544 bytes respectively. Every tier is mixed precision (different layers use different bit widths), so the nominal bit width can mislead; BPW reflects quality density per unit of size.

BPW for the two NInfer tiers is the whole .ninfer file bytes × 8 ÷ total parameter count 27,781,427,952: both files were 23,477,856,260 bytes before the vision component was added, so BPW is 6.76 for each. Each .ninfer package includes the Q8 MTP head, the DFlash2 draft, and the proposal head; the vision component added on 2026-10-09 (295,720,448 bytes) is not counted in this BPW; the safetensors-based BPW above for NVFP4 / BF16 / FP8 excludes the DFlash2 draft, so the two formulas differ and BPW should not be compared directly across those families.

INT8 W8A8's BPW counts only its 8 INT8 text weight files (the BF16 vision file model_visual.safetensors, 921,497,224 bytes, is not counted): their total bytes, 29,480,791,128, × 8 ÷ the text parameter count 26,895,998,464, giving 8.77. The text parameter count comes from the safetensors headers of the 8 INT8 weight files (INT8 and BF16 tensors count their elements; scale tensors add no elements) and equals the BF16 package's count without the vision tower (460,730,096) and MTP (424,699,392); because the denominator differs from the 27,781,427,952 used above, BPW should not be compared directly with the other tiers.

• Each tier carries its own copy of the DFlash2 draft (identical files, 2,407,384,680 bytes): NVFP4/W4A16/DFlash2-FP8/, NVFP4/W4A4/DFlash2-FP8/, and NVFP4/W4A4-W8A8/DFlash2-FP8/, byte-identical to the drafts in the main repository.
• For a BF16 DFlash2 draft (e.g. for V100 or engines that cannot load the FP8 draft), use the upstream incoai/Qwen3.8-27B-DFlash2.
• DFlash2 draft format: FP8 E4M3 weights with 128×128 block scales (weight_scale_inv), the same format as the FP8 body; q/k/v projections, fc, the candidate selector, convolutions and norms stay BF16. It replaced the earlier per-tensor-scale draft on 2026-10-08 (engines that only read block scales, such as radiance, loaded the old draft without error but accepted almost no draft tokens). Checked on one RTX PRO 6000 on 2026-10-09 (SGLang 0.5.21, FP8 body, --speculative-num-draft-tokens 8 --speculative-dflash-block-size 8, one request at a time, 6 coding prompts, 1,024-token cap): greedy mean accept length is 5.05 at 171 tok/s per request (median), against 4.99 at 178 tok/s for the earlier draft with the same command. At temperature 1.0 the two run at 155 and 152 tok/s. In SGLang the new draft accepts as well as the old one. The other DFlash2 speed numbers in this card were measured with the earlier draft.
• The three NVFP4 tiers were each benchmarked on their own package.
• NVFP4 scores were measured on a single GPU at C8; see "Scores" above for the protocol.
• NInfer tier scores are carried over from the 2026-10-05 triad on the same text path (single GPU, C8, official NInfer, no speculation); see the NInfer tiers section above. NInfer speeds were measured separately on the new packages; see "NInfer tiers" under Best-TPS and concurrency recommendations below.
• INT8 W8A8 scores are a full triad measured on a single GPU at C8 with vLLM and no speculation; see the INT8 W8A8 section above for the protocol. Its speeds come from a quick benchmark; see "INT8 W8A8" under Best-TPS and concurrency recommendations below.
• INT4 W4A16 scores are a full triad measured on a single GPU at C8 with vLLM 0.28 and MTP (3 draft tokens, built-in head); see the INT4 W4A16 section above for the protocol, and "INT4 W4A16" under Best-TPS and concurrency recommendations below for speeds.

Best-TPS and concurrency recommendations

Environment: one RTX PRO 6000 Blackwell 96GB; SGLang, 32K context (32,768), 4,096-token generation cap, thinking enabled, model-default sampling. Data from evaluation/BENCH_NVFP4_W4A16.json, evaluation/BENCH_NVFP4_W4A4.json, and evaluation/BENCH_NVFP4_MIXED.json (the W4A4-W8A8 tier).

• Which speculation: on W4A4 and W4A4-W8A8, DFlash2 beats MTP on both aggregate and per-request speed at every measured concurrency (C1–C16). On W4A16, MTP has the higher aggregate only at C4 (351.4 vs. 315.2); DFlash2 is faster at every other level and on per-request speed throughout.
• MTP only: on W4A16, use C1 for the fastest single request (128.9 tok/s), C4 for interactive use with a few users (351.4 aggregate, 108.0 per request), C8 for shared serving (450.7 aggregate, 91.3 per request), and C16 when only total throughput matters (817.3 aggregate, 70.6 per request). On W4A4 the same levels give C1 124.5 tok/s, C4 323.2 / 103.9, C8 355.5 / 86.6, and C16 912.4 / 73.0; on W4A4-W8A8, C1 116.3 tok/s, C4 213.0 / 99.8, C8 417.2 / 83.2, and C16 580.2 / 69.4 (aggregate / per request).
• Accept length: MTP 2.608–2.967 (W4A16), 2.597–2.923 (W4A4), and 2.647–2.865 (W4A4-W8A8); DFlash2 3.351–4.143 (W4A16), 3.723–4.327 (W4A4), and 3.828–4.364 (W4A4-W8A8).
• Concurrency: above C16 was not tested.

Requests per level: 4 at C1, 12 at C4, 24 at C8, 48 at C16. "vs. no spec" is the ratio of aggregate throughput at the same concurrency.

Requests per level: 4 at C1, 12 at C4, 24 at C8, 48 at C16. "vs. no spec" is the ratio of aggregate throughput at the same concurrency.

Requests per level: 4 at C1, 12 at C4, 24 at C8, 48 at C16. "vs. no spec" is the ratio of aggregate throughput at the same concurrency.

NInfer tiers

Setup: one RTX PRO 6000 Blackwell; official NInfer with the same context flags as the NInfer launch commands below (--max-context 102400 --kv-capacity 819200 --max-concurrency 8 --kv-dtype fp8); 4,096 generation cap, thinking on (template default xhigh), sampling temperature 1.0, top_p 0.95, top_k 20, same prompts as the NVFP4 benchmark above; 4 requests at C1, 12 at C4, 24 at C8. MTP uses 4 draft tokens and DFlash2 uses 8, both with --lm-head-draft. Numbers are aggregate decode tok/s. NInfer supports at most 8 concurrent requests, so C16 was not tested.

• Best setting: for both tiers and all three modes, aggregate throughput peaks at C8. NInfer W4A4 tops out with DFlash2 · C8 (975.9 tok/s) and NInfer W4A4-W8A8 with DFlash2 · C8 (804.9 tok/s); with MTP only, C8 gives 678.3 and 663.9 tok/s respectively.
• C8 versus NVFP4 W4A4 (SGLang): NInfer W4A4 reaches 306.8 vs 316.8 with no speculation, 678.3 vs 355.5 with MTP, and 975.9 vs 689.8 tok/s with DFlash2. The engines and context settings differ, so treat this as a rough comparison.

INT8 W8A8

Setup: one RTX PRO 6000 Blackwell; vLLM 0.28 with --max-model-len 102400 and --max-num-seqs 8; 512 generation cap, thinking on (xhigh), sampling temperature 1.0, top_p 0.95, top_k 20; each level sends only as many requests as its concurrency (1 at C1, 2 at C2, 4 at C4, 8 at C8). This is a quick benchmark with few requests and short outputs, so the numbers are not directly comparable with the tiers above. Numbers are aggregate tok/s (all output tokens ÷ wall-clock time). Both rows were measured on a separate, unreleased comparison build with MTP (the same INT8 text weights plus one BF16 MTP layer); with no speculation the MTP layer is not used. This package does not include MTP; a separate MTP draft folder is now provided (see below).

• Recommendation (this table): no speculation at C8 (177.3 tok/s); aggregate throughput rises with concurrency, and nothing above C8 was tested.
• The MTP row above came from an earlier, unreleased comparison build and is kept for reference only; it does not describe the MTP-BF16/ draft folder, which is measured separately below. This tier has no DFlash2 draft.

With the MTP-BF16/ draft folder

Setup: one NVIDIA RTX PRO 6000 Blackwell (96 GB); vLLM 0.28 with --max-model-len 102400, --max-num-seqs 8, and --gpu-memory-utilization 0.85; MTP row uses --speculative-config '{"method":"mtp","model":"./INT8/W8A8/MTP-BF16","num_speculative_tokens":3}'. 1024 generation cap, thinking on, sampling temperature 1.0, top_p 0.95, top_k 20; each level sends 2 × concurrency requests over the same 8 prompts. Numbers are aggregate tok/s (all output tokens ÷ wall-clock time); the last column is the median per-request decode speed at C1. Different hardware and settings from the table above, so compare rows only within this table.

• Accept length (mean tokens per step, including the verified token): 2.74–2.90 across C1–C8.
• With this draft, 3 draft tokens is faster at every level on this machine: 65.9 tok/s at C1 against 31.1 with no speculation, and 389.8 tok/s at C8 against 127.0.
• The INT8 scores above were measured without speculation; the draft folder does not change any INT8 weight file.

INT4 W4A16

Setup: one NVIDIA RTX PRO 6000 Blackwell (96 GB); vLLM 0.28 with --max-model-len 102400, --max-num-seqs 8, and --gpu-memory-utilization 0.85; MTP row uses the built-in head with --speculative-config '{"method":"mtp","num_speculative_tokens":3}'. 1024 generation cap, thinking on, sampling temperature 1.0, top_p 0.95, top_k 20; each level sends 2 × concurrency requests over the same 8 prompts (same script as the INT8 MTP-BF16/ table). Numbers are aggregate tok/s (all output tokens ÷ wall-clock time); the last column is the median per-request decode speed at C1.

• Accept length (mean tokens per step, including the verified token): 2.77–2.92 across C1–C8 in this table; 2.81 averaged over the full eval run.
• The INT4 scores above were measured with MTP (3 draft tokens).

Recommended launch commands

Quantization scheme

• NVFP4 W4A16 (NVFP4/W4A16/): ModelOpt local-Hessian calibration. NVFP4 is E2M1 in 16-element blocks with one E4M3 scale per block plus a per-tensor FP32 second-level scale. MLP gate/up/down, full-attention q/k/v/o, and linear-attention in_proj_qkv / in_proj_z / out_proj (400 Linear layers) use NVFP4 weights with BF16 activations; linear-attention in_proj_a / in_proj_b and lm_head (97 layers) stay in BF16. The quantization metadata uses ModelOpt's MIXED_PRECISION format (each of the 400 layers tagged W4A16_NVFP4), which SGLang detects as modelopt_mixed. 1,999 tensors: 1,651 language, 333 vision, 15 MTP.
• NVFP4 W4A4 (NVFP4/W4A4/): same calibration data and algorithm and the same 400 Linear layers, with both weights and activations in NVFP4 (activations are quantized at inference in 16-element blocks). Standard NVFP4 export, detected by SGLang as modelopt_fp4. 2,399 tensors: 2,051 language, 333 vision, 15 MTP.
• NVFP4 W4A4-W8A8 (NVFP4/W4A4-W8A8/): mixed precision with the same calibration data and algorithm. MLP gate/up/down (192 Linear layers) use NVFP4 W4A4 (both weights and activations in NVFP4; activations are quantized at inference in 16-element blocks); full-attention q/k/v/o and linear-attention in_proj_qkv / in_proj_z / out_proj (208 Linear layers) use FP8 W8A8 (E4M3 weights and activations with static per-tensor scales; activation scales come from calibration); linear-attention in_proj_a / in_proj_b and lm_head (97 layers) stay in BF16. The quantization metadata uses ModelOpt's MIXED_PRECISION format (each layer tagged NVFP4 or FP8), which SGLang detects as modelopt_mixed. 2,191 tensors: 1,843 language, 333 vision, 15 MTP.
• NVFP4 calibration data: complete trajectories (prompt + reasoning + final answer) that were correct and non-empty, taken from this model's own RLOO training data: 512 samples, about 1.735M tokens, at most 4,096 tokens each. No official GPQA, MMLU, or LCB problems are used. See evaluation/NVFP4_CALIBRATION_METHOD.md. In all three tiers the vision tower (333 tensors) and MTP (15 tensors) are BF16.
• Architecture: Qwen3_5ForConditionalGeneration, 64 layers (linear attention and full attention alternating 3:1).

NVFP4 smoke tests: patch-free load in SGLang, greedy parity, and the image request passed. MTP and DFlash2 speed and accept length come from the separate benchmark runs (all C1–C16 levels completed; see above).

Commands

Requires official SGLang ≥ 0.5.19; vLLM ≥ 0.28 (MTP only). vLLM MTP on NVFP4 with tensor parallel ≥ 2 has a known issue (vLLM #52480); use a single GPU there. INT8 W8A8 and INT4 W4A16 load only in vLLM (tested on 0.28); their commands are at the end of this section.

MTP and DFlash2 are mutually exclusive; enable only one per launch. The examples use ./NVFP4/W4A16; for the W4A4 or W4A4-W8A8 tier, replace --model-path with ./NVFP4/W4A4 or ./NVFP4/W4A4-W8A8. The DFlash2 draft is the DFlash2-FP8 folder inside each tier (./NVFP4/W4A16/DFlash2-FP8, ./NVFP4/W4A4/DFlash2-FP8, ./NVFP4/W4A4-W8A8/DFlash2-FP8). SGLang detects the NVFP4 quantization format automatically, so no --quantization flag is needed; NVFP4 was tested only on SGLang. Concurrency and context settings match the benchmark runs; adjust as needed.

SGLang · NVFP4 W4A16 · MTP (bundled BF16) (for W4A4 or W4A4-W8A8, replace W4A16 with that directory name)

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A16 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4

vLLM · NVFP4 · MTP (one command per tier, pick one; run on a single GPU; the NVFP4 packages in this repository were not tested on vLLM)

# W4A16
vllm serve ./NVFP4/W4A16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# W4A4
vllm serve ./NVFP4/W4A4 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# W4A4-W8A8
vllm serve ./NVFP4/W4A4-W8A8 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

SGLang · NVFP4 W4A16 · DFlash2 (for W4A4, replace both W4A16 with W4A4)

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A16 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path ./NVFP4/W4A16/DFlash2-FP8 \
  --speculative-draft-model-quantization fp8 \
  --speculative-num-draft-tokens 8

SGLang · NVFP4 W4A4-W8A8 · DFlash2

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A4-W8A8 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path ./NVFP4/W4A4-W8A8/DFlash2-FP8 \
  --speculative-draft-model-quantization fp8 \
  --speculative-num-draft-tokens 8

The NVFP4 scores in this card were measured on SGLang with DFlash2 enabled (NInfer and INT8 W8A8 without speculation, INT4 W4A16 with MTP; see each section); speculative decoding does not change the target model's output distribution. Default sampling is in generation_config.json (temperature 1.0, top_p 0.95, top_k 20).

Images on SGLang / vLLM: NVFP4/W4A16/, NVFP4/W4A4/ and NVFP4/W4A4-W8A8/ are full multimodal checkpoints: config.json declares Qwen3_5ForConditionalGeneration with a vision_config, each folder has preprocessor_config.json and video_preprocessor_config.json, and the weight index maps all 333 vision-tower tensors (to vision-bf16.safetensors in that folder). The SGLang and vLLM commands above accept images as they are, with no extra flag; send OpenAI-style image_url parts to POST /v1/chat/completions (request example under NInfer · vision below; set model to the served model name). INT8/W8A8/ accepts images too since 2026-10-09 (Qwen3_5ForConditionalGeneration; BF16 vision tower in model_visual.safetensors, preprocessor configs included): use the INT8 vLLM commands below as they are. INT4/W4A16/ accepts images too (BF16 vision tower in model_visual.safetensors): use the INT4 vLLM command below.

NVFP4 · NInfer (.ninfer, official NInfer)

W4A4 is shown; for W4A4-W8A8, change the path to ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer and --model-id to qwen3.8-27b-coder390-w4a4-w8a8.

The MTP-DFlash2 packages already contain the Q8 MTP head and the DFlash2 draft, so NInfer needs no external draft model (unlike the SGLang commands above, which take a DFlash2 draft path); pass either --spec mtp or --spec dflash2, or omit --spec to run without speculation.

Images: the two V3 MTP-DFlash2 .ninfer packages in NVFP4-NInfer/ (W4A4 and W4A4-W8A8) contain the vision component since 2026-10-09 (295,720,448 bytes, quantized from the BF16 vision weights: patch embedding Q6, attention q/k/v and MLP fc1 Q4, other projections Q5, merger Q8, norms and biases BF16). Text, MTP and DFlash2 weights are byte-for-byte unchanged. Add --vision to any NInfer command below and send images as OpenAI-style image_url parts to POST /v1/chat/completions. Without --vision the packages run text-only, as before. The V2 packages (-NInferV2-NOT-FOR-V3) are unchanged and still have no vision. GPU test on 2026-10-09 (one RTX PRO 6000, NInfer-all 64492cb with --vision, image prompt): both MTP-DFlash2 packages answered correctly. W4A4-MTP-DFlash2: MTP (3 draft tokens) accept length 2.89, decode 153 tok/s; DFlash2 (7 draft tokens) accept length 3.53, decode 185 tok/s. W4A4-W8A8-MTP-DFlash2: MTP 3.19, 164 tok/s; DFlash2 3.74, 195 tok/s.

NInfer · no speculation

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8

NInfer · MTP (4 draft tokens)

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft

NInfer · DFlash2 (8 draft tokens)

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft

NInfer · vision (image input), one command per package

Each command is the matching NInfer command above plus --vision. Nothing else is needed: the vision component is inside every V3 .ninfer, so the vision-bf16.safetensors next to the packages is not used at run time.

# W4A4 · MTP-DFlash2 package · no speculation
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --vision

# W4A4 · MTP-DFlash2 package · MTP
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft \
  --vision

# W4A4 · MTP-DFlash2 package · DFlash2
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft \
  --vision

# W4A4-W8A8 · MTP-DFlash2 package · no speculation
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --vision

# W4A4-W8A8 · MTP-DFlash2 package · MTP
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft \
  --vision

# W4A4-W8A8 · MTP-DFlash2 package · DFlash2
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft \
  --vision

Image request: OpenAI-style image_url part in POST /v1/chat/completions; url takes an http(s):// image link; a data URI of a local image also works, and model must match --model-id.

curl http://127.0.0.1:19931/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b-coder390-w4a4",
    "messages": [{"role": "user", "content": [
      {"type": "image_url", "image_url": {"url": "https://example.com/demo.png"}},
      {"type": "text", "text": "Describe this image."}
    ]}],
    "max_tokens": 2048
  }'

Health check: curl -sf http://127.0.0.1:19931/health. The chat endpoint is POST /v1/chat/completions; the request model must match --model-id.

vLLM · INT8 W8A8 (no speculation)

VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel \
VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT8/W8A8 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int8 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3

On RTX PRO 6000 Blackwell, vLLM's Cutlass INT8 kernel is unavailable and the server will not start unless it is disabled with VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel; vLLM then uses its Triton INT8 kernel (the startup log shows TritonInt8ScaledMMLinearKernel). VLLM_USE_FLASHINFER_SAMPLER=0 matches the tested setup. Tested on vLLM 0.28; SGLang cannot load this package (ModelOptFp8Config raises an error). --reasoning-parser qwen3 returns the reasoning separately in reasoning_content (it was off during scoring).

vLLM · INT8 W8A8 · MTP (external draft INT8/W8A8/MTP-BF16/)

VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel \
VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT8/W8A8 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int8 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","model":"./INT8/W8A8/MTP-BF16","num_speculative_tokens":3}'

On an RTX PRO 6000 Blackwell, set VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel and VLLM_USE_FLASHINFER_SAMPLER=0, as in the command above. Both rows of the new table used those two settings.

vLLM · INT4 W4A16 · MTP (built-in head) · text and images

VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT4/W4A16 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int4 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Images go as OpenAI-style image_url parts with the same server; no extra flag is needed:

curl -s http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "int4",
  "messages": [{"role": "user", "content": [
    {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
    {"type": "text", "text": "Describe this image."}
  ]}],
  "max_tokens": 2048
}'

Remove --speculative-config to run without speculation. Tested on vLLM 0.28 (VLLM_USE_FLASHINFER_SAMPLER=0 matches the tested setup); SGLang and NInfer were not tested with this package. --reasoning-parser qwen3 returns the reasoning separately.

Files overview

NVFP4/W4A16/ (main package: 17 files, 20,613,406,953 bytes; the DFlash2 subdirectory is listed separately below)

NVFP4/W4A16/SHA256SUMS covers the first 16 files below; its own SHA256 is ecc28d31be0d97e23509af61625251af14980db5b67189a1d1ca75a3fc9652f3. mtp-bf16.safetensors is byte-identical to the file of the same name in each tier of the main repository.

NVFP4/W4A16/DFlash2-FP8/ (4 files, 2,407,384,680 bytes)

Byte-identical to FP8/DFlash2-FP8/ in the main repository.

NVFP4/W4A4/ (main package: 17 files, 20,613,386,462 bytes; the DFlash2 subdirectory is listed separately below)

NVFP4/W4A4/SHA256SUMS covers the first 16 files below; its own SHA256 is 5aea8bb9c1f1979661bfa5c1216de7aafeb64cd708aa8cc71f3b185c224cfde0. mtp-bf16.safetensors is byte-identical to the file of the same name in each tier of the main repository.

NVFP4/W4A4/DFlash2-FP8/ (4 files, 2,407,384,680 bytes)

Byte-identical to FP8/DFlash2-FP8/ in the main repository.

NVFP4/W4A4-W8A8/ (main package: 18 files, 23,769,634,648 bytes; the DFlash2 subdirectory is listed separately below)

NVFP4/W4A4-W8A8/SHA256SUMS covers the first 17 files below; its own SHA256 is c177f4907a7dd037efe91b8d34927dd3a3624227f259238b82db0f12f8b0b8e2. mtp-bf16.safetensors is byte-identical to the file of the same name in each tier of the main repository.

NVFP4/W4A4-W8A8/DFlash2-FP8/ (4 files, 2,407,384,680 bytes)

Byte-identical to FP8/DFlash2-FP8/ in the main repository.

evaluation/

Score, benchmark, and calibration records for the NVFP4 W4A16, W4A4, and W4A4-W8A8 tiers; evaluation/SHA256SUMS is the checksum list for the directory. NVFP4_*.results.summary.json are the raw summaries of the three suites for the three NVFP4 tiers (NVFP4_MIXED_* is the W4A4-W8A8 tier).

Verification

(cd NVFP4/W4A16 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4/W4A4 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4/W4A4-W8A8 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4-NInfer/W4A4 && sha256sum -c SHA256SUMS)
(cd NVFP4-NInfer/W4A4-W8A8 && sha256sum -c SHA256SUMS)
(cd INT8/W8A8 && sha256sum -c SHA256SUMS && cd MTP-BF16 && sha256sum -c SHA256SUMS)
(cd INT4/W4A16 && sha256sum -c SHA256SUMS)
(cd evaluation && sha256sum -c SHA256SUMS)

NInfer loading notes

• Only the NInfer engine can load .ninfer; SGLang, vLLM, llama.cpp, and transformers cannot.
• Use ninfer-serve from official NInfer (Neroued/ninfer, master branch); no patches are needed. (Only the GGUF-derived .ninfer packages need NInfer-all.) Tested on one RTX PRO 6000 Blackwell (compute capability 12.0); NInfer supports at most 8 concurrent requests.
• NVFP4-NInfer/W4A4/ and NVFP4-NInfer/W4A4-W8A8/ each hold one .ninfer file with MTP and DFlash2: the text path is the same one scored on 2026-10-05, and the package also carries a Q8 MTP head, the DFlash2 draft, the proposal head, and (since 2026-10-09) the vision component, so --vision enables image input; the vision-bf16.safetensors in the same folder is the BF16 source of that part and is not needed at run time. The MTP projections are Q8 because several of their matrix shapes have no BF16 kernel in NInfer, and a BF16 MTP makes the server exit while building the compute graph.
• MTP and DFlash2 are mutually exclusive per launch (--spec mtp or --spec dflash2); --lm-head-draft must be used together with --spec. MTP accepts 1–5 draft tokens and DFlash2 1–15; the NInfer commands under Commands above use 4 and 8, as in the benchmark. The DFlash2 draft is built into the package, so no separate DFlash2-FP8/ directory is needed.
• These context flags are verified to start: --max-context 102400 --kv-capacity 819200 --max-concurrency 8 --kv-dtype fp8. --kv-capacity must fall between max-context and max-context × max-concurrency; 819,200 is exactly 8 × 102,400.
• The HF packages under the NVFP4 directories still run in SGLang (MTP or DFlash2); they are not interchangeable with the NInfer packages.

NInfer V2 MTP packages (vision, no DFlash2, Blackwell)

Two kinds of V2 package; do not mix them up (neither runs on the current V3 engines, i.e. current official NInfer and NInfer-all):

Usage for the V100 packages is in the next section and is unchanged.

Added 2026-10-11: two NInfer V2 container packages for Blackwell cards that need a smaller file (for example 32 GB cards). They contain the NVFP4 text weights, the Q8 MTP head and the vision component, and no DFlash2 draft. Each file is about 2.3 GB smaller than the V3 MTP-DFlash2 package.

• Engine: official NInfer before its switch to V3, commit 99f08f00 (2026-09-13). It builds and runs only on Blackwell (sm_120) and needs CUDA 13.1 or newer. Current official NInfer and NInfer-all read V3 only and reject V2 packages; the V100 ninfer-v100 branch is not a target for these two packages either.
• Features: --spec mtp and --vision work; --spec dflash2 does not.
• W4A4 vs W4A4-W8A8: the source quantizations differ, but after conversion the in-package format is identical (same tensors, formats and sizes; only the values differ).
• The chat template inside is the same as this repo's BF16/NVFP4 chat_template.jinja (that NInfer version accepts only this template).

Build (Blackwell, CUDA 13.1+):

git clone https://github.com/Neroued/ninfer.git && cd ninfer && git checkout 99f08f00
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=120a
cmake --build build --target ninfer-serve

Tested launch command (MTP + images):

./build/apps/ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-NVFP4-W4A4-MTP-NInferV2.ninfer \
  --host 127.0.0.1 --port 8000 \
  --max-context 102400 --kv-capacity 102400 --max-concurrency 1 --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft --vision

Drop --vision if you do not need images; that saves about 0.9 GiB of VRAM.

Measured (2026-10-11, one RTX PRO 6000 Blackwell, the command above, C1 with 4 coding requests in sequence, generation limit 4,096, thinking on, temperature 1.0, top_p 0.95, top_k 20):

Without speculative decoding, W4A4 runs at 69 tok/s under the same conditions.

VRAM (W4A4, nvidia-smi reading including the CUDA context; FP8 KV, one concurrent request, --kv-capacity equal to the context length, KV allocated once at startup; each setting also served a long prompt of about 30K / 61K / 97K tokens):

These VRAM figures were measured with --draft-tokens 3; with --draft-tokens 4 --lm-head-draft at 100K it is 25,641 MiB. The server also pins about 9 GiB of host RAM by default (8 GiB host KV cache and other state).

NInfer V2 packages (Tesla V100 only)

The four packages whose filenames contain NInferV2-NOT-FOR-V3 are for Tesla V100 only (ninfer-v100 branch); on any other GPU, use the packages without this suffix.

• The 4 packages whose filenames contain NInferV2-NOT-FOR-V3 use the NInfer V2 container and are only for the Tesla V100 ninfer-v100 branch. Official NInfer (ninfer-serve) reads V3 only and rejects these files; on other GPUs, use the V3 packages above (without NInferV2 in the name). Do not rename the V2 packages to V3 filenames or use them to replace the V3 packages.
• The two text-only packages (Qwen3.8-27B-Coder390-W4A4-NInferV2-NOT-FOR-V3.ninfer, Qwen3.8-27B-Coder390-W4A4-W8A8-NInferV2-NOT-FOR-V3.ninfer) have no MTP or DFlash2, so do not pass --spec; the two MTP-DFlash2 packages (…-MTP-DFlash2-NInferV2-NOT-FOR-V3.ninfer) accept --spec mtp or --spec dflash2; they already contain the MTP head and the DFlash2 draft, so no external draft model is needed. None of the four V2 packages includes the vision component, so they cannot take images (no --vision); for image input use the V3 packages.
• V100 has no NVFP4 KV or K8V4 KV, so do not use the V3 flags --kv-dtype rk8v4 or --kv-capacity 819200; when serving on V100, add --prefill-chunk 2048.
• The V2 packages have not been separately evaluated; they have only passed SHA256 verification. In SHA256SUMS, the two V2 lines come after a # comment line, separate from the V3 lines.

# V100 ninfer-v100 branch · text-only package (no --spec)
ninfer ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-W4A4-NInferV2-NOT-FOR-V3.ninfer \
  --prompt "Explain prefill and decode in three sentences." \
  --max-context 8192 \
  --max-new 256

# V100 ninfer-v100 branch · MTP-DFlash2 package (MTP on)
ninfer ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-W4A4-MTP-DFlash2-NInferV2-NOT-FOR-V3.ninfer \
  --prompt "Explain prefill and decode in three sentences." \
  --max-context 8192 \
  --max-new 256 \
  --spec mtp --draft-tokens 3

NVFP4-NInfer/W4A4/ (5 files, 68,925,700,831 bytes)

NVFP4-NInfer/W4A4-W8A8/ (5 files, 68,925,700,846 bytes)

INT8 W8A8 notes

• Text and image (since 2026-10-09): the architecture is Qwen3_5ForConditionalGeneration (64 text layers, linear and full attention alternating 3:1). The BF16 vision tower from the BF16 tier is a separate file (model_visual.safetensors, 333 tensors, 921,497,224 bytes), with preprocessor_config.json / video_preprocessor_config.json; it is excluded from quantization. The 8 INT8 weight files are unchanged. The package has no embedded MTP head (use MTP-BF16/) and no DFlash2 draft. Checked on one RTX PRO 6000 on 2026-10-09 (vLLM 0.28, Cutlass INT8 kernel disabled, MTP-BF16/ with 3 draft tokens): the image prompt was answered correctly, text output was normal, accept length 3.21, single-request decode 86 tok/s.
• Quantization: SmoothQuant + INT8 W8A8. Weights are quantized to INT8 per output channel, and activations are quantized to INT8 per token at runtime; the SmoothQuant pre-scale is folded into the weights, so no inference-engine patch is needed. Quantization metadata uses the compressed-tensors (int-quantized) format.
• Layer layout: MLP gate/up/down, full-attention q/k/v/o, and linear-attention in_proj_qkv / in_proj_z / out_proj make up 400 INT8 Linear layers; linear-attention in_proj_a / in_proj_b and lm_head, 97 in total, stay BF16; linear-attention control branches such as A_log, dt_bias, norm, and conv1d also stay BF16.
• Calibration and error: 512 calibration samples; held-out NLL rises from 0.50535 (BF16) to 0.52812, a perplexity ratio of 1.023 (measured before the pre-scale was folded into the weights).
• Loading: validated only on vLLM 0.28; not for SGLang or NInfer. VLLM_LOAD.json records the load requirements. On RTX PRO 6000 Blackwell the Cutlass INT8 kernel must be disabled; see Commands above.
• MTP draft: INT8/W8A8/MTP-BF16/ holds the BF16 MTP tensors only. Checked on one RTX PRO 6000 on 2026-10-09 (vLLM 0.28, Cutlass INT8 kernel disabled, num_speculative_tokens=3): mean accept length 2.74–2.90.

INT8/W8A8/ (19 files, 30,422,466,299 bytes)

INT8/W8A8/MTP-BF16/ (3 files, 849,403,270 bytes)

INT8/W8A8/MTP-BF16/SHA256SUMS covers the first 2 files below; its own SHA256 is 313b058349deba83f05adb44b79f2f76ae668337834086f8dbcfacba131bf403.

INT4 W4A16 notes

• Text and image: the architecture is Qwen3_5ForConditionalGeneration (64 text layers, linear and full attention alternating 3:1). The BF16 vision tower is a separate file (model_visual.safetensors, 333 tensors, 921,497,224 bytes, the same file as in INT8/W8A8/), with preprocessor_config.json / video_preprocessor_config.json. The built-in MTP head is BF16 (model_mtp.safetensors, 15 tensors). No DFlash2 draft.
• Quantization: GPTQ W4A16 with llm-compressor, INT4 weights, symmetric, group size 128, static activation order, dampening 0.01; activations stay BF16. lm_head, the token embedding, the MTP head and the vision tower are not quantized. Calibration: 512 sequences × 4,096 tokens rebuilt from this model's RL training rollouts. Quantization metadata uses the compressed-tensors (pack-quantized) format; recipe.yaml is the llm-compressor recipe.
• BPW 5.25: model.safetensors (17,646,891,544 bytes, text weights only) × 8 ÷ the text parameter count 26,895,998,464; the MTP and vision files are not counted.
• Image + MTP check (2026-10-09, one RTX PRO 6000, vLLM 0.28, 3 draft tokens): the image prompt was answered correctly with thinking on and off.
• Loading: validated only on vLLM 0.28. VLLM_LOAD.json records the load settings.

INT4/W4A16/ (14 files, 19,437,985,972 bytes)

Known limitations

• The long tail is not eliminated: accuracy is clearly lower for reasoning ≥48K (LCB ≥48K is 4/10 for W4A16, 7/11 for W4A4, and 4/9 for W4A4-W8A8), and 4 (W4A16), 3 (W4A4), and 2 (W4A4-W8A8) LCB problems still hit the 94K cap.
• Scores depend on the protocol above (100K context, 94,208-token cap, no timeout) and should not be compared directly with numbers from other protocols; MMLU shows sampling variation.
• NVFP4 performance was measured on a single RTX PRO 6000 Blackwell with SGLang, each tier on its own package; NVFP4 was not tested on vLLM; DFlash2 was validated only on SGLang. The NInfer tiers were measured on NInfer, and INT8 W8A8 and INT4 W4A16 on vLLM 0.28.
• NVFP4 GPQA (170, 167, and 172) is below the 178 of BF16 / FP8 in the main repository, MMLU (439, 440, and 442) is slightly below 445–450, and LCB (91, 92, and 88) is around 90; the calibration data comes from this model's own RLOO training data and contains no official evaluation problems. NVFP4 smoke tests were run only on SGLang (patch-free load, greedy parity, image request); MTP and DFlash2 are covered by the full benchmark runs.
• INT8 W8A8 runs only in vLLM (its scores were measured on text; image input was added on 2026-10-09); GPQA 175 is slightly below the 177 of the original Qwen3.8 27B (FP8); speeds come from a quick benchmark (requests per level equal to the concurrency, 512 generation cap) and are not directly comparable with the other tiers; GPQA / MMLU reasoning lengths were recounted from the raw text with the BF16 tokenizer; LCB kept no reasoning text, so it has no reasoning-length percentiles.
• Verify outputs independently, especially in high-stakes settings.

License and acknowledgements

Licensed under Apache-2.0, following the license of the base model Qwen3.8-27B.

Thanks to the Qwen team for the base model and the official FP8 scheme; to Opus5.5, GPT6Astra, Grok4.7, DSV4Pro, and K3 for problem writing, gold labels, teacher trajectories, data review, and the RLOO value review; and to the SGLang, vLLM, and DFlash communities.

Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-NVFP4-NInfer

在 Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2(官方 Qwen3.8-27B 经 SFT + SimPO 后训练得到的模型)基础上继续做多轮 SFT 与 RLOO 后训练,专治原版「已有答案却停不下来、写满 94K 也不给最终答案」的问题。本仓库提供 NVFP4 W4A16(NVFP4/W4A16/)、NVFP4 W4A4(NVFP4/W4A4/)、NVFP4 W4A4-W8A8(NVFP4/W4A4-W8A8/,混合精度)三档(完整多模态、内置官方 BF16 MTP、每档自带 DFlash2 草稿(DFlash2-FP8/),SGLang 免补丁加载),以及 NInfer W4A4(NVFP4-NInfer/W4A4/)与 NInfer W4A4-W8A8(NVFP4-NInfer/W4A4-W8A8/)两档官方 NInfer 包(每个 .ninfer 包内含 Q8 MTP 与 DFlash2 草稿,仅 NInfer 引擎可加载),INT8 W8A8(INT8/W8A8/,文本 + 图像,仅 vLLM 可加载)以及 INT4 W4A16(INT4/W4A16/,文本 + 图像,内置 BF16 MTP 头,仅 vLLM 可加载);BF16 与静态 FP8 档见主仓 Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2。

• 94K 截断大幅减少:同口径下 GPQA 从 4 降到 1,LCB 从 13 降到 3。
• 三项全面不低于原版:GPQA 178/198、MMLU 450/500、LCB 90/100(静态 FP8),原版 Qwen3.8 27B(FP8)为 177 / 444 / 83。
• 量化无损:BF16、动态 FP8、静态 FP8 三种精度 GPQA、LCB 一题不差。
• 两种投机解码(互斥):DFlash2 单请求约 180 tok/s,MTP 约 91 tok/s,不开投机约 46 tok/s。

相关仓库

• 主仓（BF16 / FP8 / NVFP4 / INT8 / GGUF 等全部版本）：Hugging Face · ModelScope
• GGUF / GGUF-NInfer（llama.cpp、NInfer，内置 MTP）：Hugging Face · ModelScope

训练方法

谱系:原版 Qwen3.8 27B(FP8;177 / 444 / 83)→ 我们的 SFT + SimPO110 终版,即 EfficientThink 基座(171 / 442 / 89)→ K3 续训 SFT → 第二周 SFT(week2dose,update-225)→ 第一轮 RLOO(182 组)→ 再一轮 SFT(merge-sft-100),得 sft-base-rloo(177 / 448 / 90)→ 第二轮 RLOO(172 组)→ Coder390(178 / 445 / 90,动态 FP8 口径)。括号内依次为 GPQA / MMLU / LCB,均为 100K 同口径全量。

训练基座:本模型以我们自己的 EfficientThink SFT + SimPO 模型(SimPO110 终版)为起点,继续做多轮 SFT 与 RLOO,该基座模型本身是官方 Qwen3.8-27B 的后训练版本,见 Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2。

要解决的问题:模型在局部已经有答案之后,仍用 Wait / Actually 把同一套推导反复再推,直到写满上下文,最终答案为空。这类题多数不是不会,而是停不下来。

RLOO 数据:每题采样 8 条轨迹,只保留组内有对有错的题(8 条全对的不入训);组内 0 或 1 条对的,补 1 条审过的短教师轨迹,替换该组最短的一条错轨迹(共 27 组)。最终 172 组、1,376 条:正确 859 条、错误 517 条,其中空答 68 条。

奖励:惩罚只打在错的一侧,长而正确的推理仍得正奖励,以免压制必要的长推理。

结果:同口径下,GPQA 的 94K 截断从原版 Qwen3.8 27B(FP8)的 4 降到 1,LCB 从 13 降到 3,三项分数均不低于原版(均为 FP8 档;NVFP4 三档见下文成绩)。

成绩

口径:同一份 Coder390 合并权重的 NVFP4 量化;单张 RTX PRO 6000,C8,上下文 100K,生成上限 94,208,客户端不计超时,DFlash2 草稿,SGLang;GPQA 198、MMLU 500、LCB 100 全量;空答记错,截断但答对记对,失败样本全部留在分母里。

LCB 按每题的 pass 字段判对错。原始记录见 evaluation/SCORES_NVFP4_W4A16.json、evaluation/SCORES_NVFP4_W4A4.json、evaluation/SCORES_NVFP4_MIXED.json(W4A4-W8A8 档)。Coder390 静态 FP8 一列取自主仓成绩(两张 RTX PRO 6000、每卡 C8,原始记录为主仓 evaluation/SCORES_STATICFP8.json),仅供对照。

NVFP4 三档:分数、思考长度、94K 截断与空答

思考长度取 usage.reasoning_tokens;94K 截断指写满生成上限。LCB 空答指没有提取到可运行代码的题数。原版 Qwen3.8 27B(FP8)与 Coder390 静态 FP8 为两张 RTX PRO 6000、每卡 C8(取自主仓成绩,仅供对照);NVFP4 三档为单张 RTX PRO 6000、C8,其余口径相同。

加粗表示优于原版 Qwen3.8 27B(FP8):分数更高,或 P50 / P70 / P90、94K 截断、空答更低;等于或不如原版的不加粗。

LCB 改善最明显:原版 Qwen3.8 27B(FP8)的 13 条 94K 截断全是空代码(截断 13、空答 13),Coder390 静态 FP8 为 3 条,NVFP4 三档降到 2–4 条。W4A4 的 3 条是第 0、53、96 题(成绩文件中的题目编号),同样是截断后没有代码;W4A16 的 4 条中,第 6、7、54 题没有代码,第 53 题有 20 个字符的代码但答错;W4A4-W8A8 的 2 条中,第 11 题没有代码,第 6 题有 2,338 个字符的代码但答错。GPQA 的 94K 截断原版 4 条,Coder390 静态 FP8 1 条,NVFP4 三档各 2 条;MMLU 原版 1 条,Coder390 静态 FP8 1 条,三档 0–2 条,没有明显变化。

分桶为「答对/该桶题数」,按思考长度左闭右开。

NInfer 两档:分数、思考长度、94K 截断与空答

口径:分数沿用 2026-10-05 在同一条 NInfer 文本路径上测得的官方三联;单张 RTX PRO 6000,官方 NInfer 引擎(Neroued/ninfer),C8,上下文 100K,生成上限 94,208,客户端不计超时,不开投机;GPQA 198、MMLU 500、LCB 100 全量;空答记错,截断但答对记对,失败样本全部留在分母里。本次上传的新包只在这条文本路径上加入 Q8 MTP、DFlash2 草稿与 proposal head,文本路径不变,全量三联没有重跑。只有 NInfer 引擎能加载 .ninfer;SGLang / vLLM / llama.cpp / transformers 均不能加载。

思考长度取 usage.reasoning_tokens(即 completion_tokens_details.reasoning_tokens);94K 截断指写满生成上限。LCB 空答指没有提取到可运行代码的题数。加粗表示优于原版 Qwen3.8 27B(FP8)。

NInfer W4A4 的 LCB 空答 2 条是第 11、15 题,均为截断后没有代码;GPQA 空答 2 条是第 79、127 题,均为截断且 pred 为空。MMLU 无截断、无空答。NInfer W4A4-W8A8 的 LCB 空答 3 条是第 17、53、54 题(第 17 题正常停但没有代码,第 53、54 题截断且没有代码);GPQA 空答 2 条是第 79、127 题,均为截断且 pred 为空。MMLU 无截断、无空答。

分桶为「答对/该桶题数」,按思考长度左闭右开。

INT8 W8A8:分数、94K 截断与空答

口径:单张 RTX PRO 6000,vLLM 0.28,C8,上下文 100K,生成上限 94,208,客户端不计超时,不开投机,思考档位 xhigh;GPQA 198、MMLU 500、LCB 100 全量;空答记错,截断但答对记对,失败样本全部留在分母里。INT8 W8A8 只能用 vLLM 加载。

94K 截断指写满生成上限。LCB 空答指没有提取到可运行代码的题数。加粗表示优于原版 Qwen3.8 27B(FP8),其数字见上方总表。

注:INT8 W8A8 测评时 vLLM 未开推理解析器,usage.reasoning_tokens 均为 0,思考与答案都记在正文里。GPQA 与 MMLU 的思考长度按原文重新计数:取正文中 </think> 之前的文本,用 BF16 tokenizer 计数(与 GGUF 档相同的方法)。LCB 的结果文件只保存了代码,没有保存思考原文,无法计数,记「—」。

GPQA 的 1 条截断是第 127 题,pred 为空,同时记为空答。LCB 的 2 条截断是第 7、53 题:第 7 题截断后留下的代码运行出错,判错;第 53 题虽然截断,代码仍判对;LCB 没有空答。MMLU 无截断、无空答。LCB 按难度为 hard 38/46、medium 30/31、easy 23/23。

INT4 W4A16:分数、94K 截断与空答

口径:单张 RTX PRO 6000,vLLM 0.28,C8,上下文 100K,生成上限 94,208,客户端不计超时,思考强度 xhigh,开 MTP 投机(内置头,3 个草稿 token),--reasoning-parser qwen3;GPQA 198 / MMLU 500 / LCB 100 全量;空答算错,截断但答对算对,所有失败都留在分母里。测评在 2026-10-09 暂停过,之后用同样的服务参数继续,已完成的题保留。

94K 截断指触到生成上限。LCB 空答指提取不到可运行代码的题数。加粗表示优于原版 Qwen3.8 27B(FP8),其数据见上面的全档表。

思考长度取自 vLLM 返回的 usage.reasoning_tokens。GPQA:截断 1 题(第 79 题),空答 1 题(第 79 题)。MMLU 没有截断,也没有空答。LCB:截断 2 题(第 atcoder:arc182_e:6, atcoder:arc195_d:54 题),空答 2 题(第 atcoder:arc182_e:6, atcoder:arc195_d:54 题)。LCB 按难度:难 37/46,中 27/31,易 23/23。整轮测评的 MTP 平均接受长度(vLLM 日志,3 个草稿 token):2.81。

量化档

BPW(bits per weight)是整包的有效平均位宽:整包主模型权重文件(.safetensors,含视觉塔、MTP 与 scale 张量,不含 DFlash2 草稿)总字节数 × 8 ÷ 总参数数。总参数数按 safetensors 头部统计,三档一致,均为 27,781,427,952 个(NVFP4 的 U8 打包张量一个字节算 2 个元素,scale 张量不计元素)。NVFP4 W4A16、NVFP4 W4A4、NVFP4 W4A4-W8A8 的权重文件总字节数依次为 20,593,097,872、20,593,145,056、23,749,329,544。各档都是混合精度(不同层不同位宽),标称位宽会误导,BPW 反映单位体积下的质量密度。

NInfer 两档的 BPW 按整包 .ninfer 文件字节数 × 8 ÷ 总参数数 27,781,427,952 计算:两档文件加视觉前均为 23,477,856,260 字节,BPW 均为 6.76。.ninfer 包内含 Q8 MTP、DFlash2 草稿与 proposal head;2026-10-09 加入的视觉部分(295,720,448 字节)不计入这里的 BPW;上方 NVFP4 / BF16 / FP8 的 safetensors 计法不计 DFlash2 草稿,两种计法不同,不宜直接横向比较 BPW。

INT8 W8A8 的 BPW 只算 8 个 INT8 文本权重文件(BF16 视觉文件 model_visual.safetensors 921,497,224 字节不计入),按其总字节数 29,480,791,128 × 8 ÷ 文本参数数 26,895,998,464 计算,为 8.77。文本参数数按 8 个 INT8 权重文件的 safetensors 头部统计(INT8 与 BF16 张量计元素,scale 张量不计元素),与 BF16 包去掉视觉塔(460,730,096)和 MTP(424,699,392)后的参数数一致;分母与上方各档的 27,781,427,952 不同,不宜直接横向比较 BPW。

• 每档各自带一份 DFlash2 草稿(文件完全相同,2,407,384,680 字节):NVFP4/W4A16/DFlash2-FP8/、NVFP4/W4A4/DFlash2-FP8/、NVFP4/W4A4-W8A8/DFlash2-FP8/,与主仓各档的草稿逐字节相同。
• BF16 版 DFlash2 草稿(如 V100 或无法加载 FP8 草稿的引擎)请用上游 incoai/Qwen3.8-27B-DFlash2。
• DFlash2 草稿格式:FP8 E4M3 权重,128×128 分块 scale(weight_scale_inv),与 FP8 主体相同;q/k/v 投影、fc、候选选择器、卷积和 norm 保持 BF16。2026-10-08 起替换了原来按张量 scale 的草稿(只认分块 scale 的引擎,如 radiance,加载旧草稿不报错,但几乎不接受草稿 token)。2026-10-09 在一台 RTX PRO 6000 上用 SGLang 0.5.21 实测新草稿:FP8 主体,--speculative-num-draft-tokens 8 --speculative-dflash-block-size 8,单请求,6 条编程提示词,生成上限 1,024。贪心平均接受长度 5.05、单请求 171 tok/s(中位数);同一命令下旧草稿是 4.99、178 tok/s。temperature 1.0 时两者分别是 155 和 152 tok/s。在 SGLang 中新草稿的接受效果与旧草稿相当。本卡其他 DFlash2 测速数字是用旧草稿测的。
• NVFP4 三档各自单独测速。
• NVFP4 三档的分数为单卡 C8 测得,口径见上文「成绩」。
• NInfer 两档成绩沿用 2026-10-05 在同一文本路径上以单卡 C8、官方 NInfer、不开投机测得的三联,口径见上文「NInfer 两档」;NInfer 两档的速度为新包单独实测,见下文「最佳 TPS 推荐与并发推荐」中的「NInfer 两档」。
• INT8 W8A8 成绩为单卡 C8、vLLM、不开投机测得的全量三联,口径见上文「INT8 W8A8」;速度为快测,见下文「最佳 TPS 推荐与并发推荐」中的「INT8 W8A8」。
• INT4 W4A16 成绩为单卡 C8、vLLM 0.28、开 MTP(草稿长度 3,内置头)测得的全量三联,口径见上文「INT4 W4A16」;速度见下文「最佳 TPS 推荐与并发推荐」中的「INT4 W4A16」。

最佳 TPS 推荐与并发推荐

测试环境:单张 RTX PRO 6000 Blackwell 96GB;SGLang,上下文 32K(32,768),生成上限 4,096,开思考,采样用模型默认值。数据取自 evaluation/BENCH_NVFP4_W4A16.json、evaluation/BENCH_NVFP4_W4A4.json、evaluation/BENCH_NVFP4_MIXED.json(W4A4-W8A8 档)。

• 选哪种投机:W4A4 与 W4A4-W8A8 在 C1–C16 每一档,DFlash2 的总吞吐和单请求速度都高于 MTP;W4A16 只有 C4 的总吞吐是 MTP 更高(351.4 对 315.2),其余各档与全部单请求速度都是 DFlash2 更高。
• 只用 MTP 时:W4A16 追求单请求速度用 C1(128.9 tok/s),交互与少量并发可用 C4(总吞吐 351.4,单请求 108.0),多人共用用 C8(总吞吐 450.7,单请求 91.3),只看总量用 C16(总吞吐 817.3,单请求 70.6);W4A4 依次为 C1(124.5 tok/s)、C4(总吞吐 323.2,单请求 103.9)、C8(总吞吐 355.5,单请求 86.6)、C16(总吞吐 912.4,单请求 73.0);W4A4-W8A8 依次为 C1(116.3 tok/s)、C4(总吞吐 213.0,单请求 99.8)、C8(总吞吐 417.2,单请求 83.2)、C16(总吞吐 580.2,单请求 69.4)。
• 接受长度:MTP 为 2.608–2.967(W4A16)、2.597–2.923(W4A4)与 2.647–2.865(W4A4-W8A8);DFlash2 为 3.351–4.143(W4A16)、3.723–4.327(W4A4)与 3.828–4.364(W4A4-W8A8)。
• 并发:C16 以上未测。

每档请求数:C1 4 个、C4 12 个、C8 24 个、C16 48 个。「相对不开投机」为同并发下总吞吐之比。

每档请求数:C1 4 个、C4 12 个、C8 24 个、C16 48 个。「相对不开投机」为同并发下总吞吐之比。

每档请求数:C1 4 个、C4 12 个、C8 24 个、C16 48 个。「相对不开投机」为同并发下总吞吐之比。

NInfer 两档

测试环境:单张 RTX PRO 6000 Blackwell;官方 NInfer,上下文参数与下文 NInfer 启动命令相同(--max-context 102400 --kv-capacity 819200 --max-concurrency 8 --kv-dtype fp8);生成上限 4,096,开思考(模板默认 xhigh),采样 temperature 1.0、top_p 0.95、top_k 20,提示词与上方 NVFP4 测速相同;C1 发 4 个请求、C4 发 12 个、C8 发 24 个。MTP 草稿长度 4、DFlash2 草稿长度 8,均开 --lm-head-draft。数字为合计解码 tok/s。NInfer 并发上限为 8,C16 未测。

• 最佳组合:两档、三种模式的总吞吐都在 C8 最高。NInfer W4A4 最高为 DFlash2 · C8(975.9 tok/s),NInfer W4A4-W8A8 最高为 DFlash2 · C8(804.9 tok/s);只用 MTP 时 C8 分别为 678.3 与 663.9 tok/s。
• 与 NVFP4 W4A4(SGLang)的 C8 对比:NInfer W4A4 不开投机 306.8 对 316.8,MTP 678.3 对 355.5,DFlash2 975.9 对 689.8 tok/s。两者引擎与上下文设置不同,仅作参考。

INT8 W8A8

测试环境:单张 RTX PRO 6000 Blackwell;vLLM 0.28,--max-model-len 102400、--max-num-seqs 8;生成上限 512,开思考(xhigh),采样 temperature 1.0、top_p 0.95、top_k 20;每档只发与并发数相同的请求数(C1 发 1 个、C2 发 2 个、C4 发 4 个、C8 发 8 个)。这是快测:请求少、生成短,数字不能与上面各档直接比较。数字为总吞吐 tok/s(全部输出 token ÷ 墙钟时间)。两行都在另做的内置 MTP 对照版上测得(同一份 INT8 文本权重另加一层 BF16 MTP,未发布);不开投机时 MTP 层不参与计算。本包不含 MTP;现另提供单独的 MTP 草稿目录(见下文)。

• 推荐(本表):不开投机,C8(177.3 tok/s);并发越高总吞吐越高,C8 以上未测。
• 上表 MTP 一行来自早先未发布的对照版,仅作参考,不代表 MTP-BF16/ 草稿目录;该目录的结果见下文。本档没有 DFlash2 草稿。

使用 MTP-BF16/ 草稿目录

测试环境:一台 NVIDIA RTX PRO 6000 Blackwell(96 GB);vLLM 0.28,--max-model-len 102400、--max-num-seqs 8、--gpu-memory-utilization 0.85;MTP 一行用 --speculative-config '{"method":"mtp","model":"./INT8/W8A8/MTP-BF16","num_speculative_tokens":3}'。生成上限 1024,开思考,采样 temperature 1.0、top_p 0.95、top_k 20;每档发 2 × 并发数个请求,轮流使用同一组 8 条提示词。数字为总吞吐 tok/s(全部输出 token ÷ 墙钟时间),最后一列为 C1 单请求解码速度中位数。硬件和设置都与上表不同,只在本表内比较。

• 接受长度(每步平均产出 token 数,含验证 token):C1–C8 为 2.74–2.90。
• 用这份草稿、3 个草稿 token,这台机器上每一档都更快:C1 为 65.9 tok/s,不开投机是 31.1;C8 为 389.8 tok/s,不开投机是 127.0。
• 上面的 INT8 分数是不开投机测的;草稿目录不改动任何 INT8 权重文件。

INT4 W4A16

环境:单张 NVIDIA RTX PRO 6000 Blackwell(96 GB);vLLM 0.28,--max-model-len 102400、--max-num-seqs 8、--gpu-memory-utilization 0.85;MTP 一行用内置头,--speculative-config '{"method":"mtp","num_speculative_tokens":3}'。生成上限 1024,开思考,采样温度 1.0、top_p 0.95、top_k 20;每档并发发送 2 × 并发数个请求,共用 8 条提示(与 INT8 MTP-BF16/ 表同一脚本)。数值为总吞吐 tok/s(全部输出 token ÷ 墙钟时间);最后一列是 C1 单请求 decode 速度中位数。

• 接受长度(每步平均产出 token 数,含校验 token):本表 C1–C8 为 2.77–2.92;整轮测评平均 2.81。
• 上面的 INT4 分数是开 MTP(3 个草稿 token)测得的。

推荐启动脚本

量化方案

• NVFP4 W4A16(NVFP4/W4A16/):ModelOpt local-Hessian 校准;NVFP4 为 E2M1,16 个元素一块,每块一个 E4M3 尺度,另加张量级 FP32 二级尺度。MLP gate/up/down、全注意力 q/k/v/o、线性注意力 in_proj_qkv / in_proj_z / out_proj 共 400 个 Linear 为 NVFP4 权重、BF16 激活;线性注意力 in_proj_a / in_proj_b 与 lm_head 共 97 个保持 BF16。量化元数据为 ModelOpt MIXED_PRECISION 格式(400 层逐层标 W4A16_NVFP4),SGLang 自动识别为 modelopt_mixed。张量共 1,999 个:语言模型 1,651、视觉 333、MTP 15。
• NVFP4 W4A4(NVFP4/W4A4/):同一份校准数据和算法,同样 400 个 Linear,权重与激活都为 NVFP4(激活在推理时按 16 个元素一块量化);标准 NVFP4 导出,SGLang 自动识别为 modelopt_fp4。张量共 2,399 个:语言模型 2,051、视觉 333、MTP 15。
• NVFP4 W4A4-W8A8(NVFP4/W4A4-W8A8/):混合精度,同一份校准数据和算法。MLP gate/up/down 共 192 个 Linear 为 NVFP4 W4A4(权重与激活都为 NVFP4,激活在推理时按 16 个元素一块量化);全注意力 q/k/v/o 与线性注意力 in_proj_qkv / in_proj_z / out_proj 共 208 个 Linear 为 FP8 W8A8(E4M3 权重与激活,每张量静态 scale,激活 scale 由校准统计得到);线性注意力 in_proj_a / in_proj_b 与 lm_head 共 97 个保持 BF16。量化元数据为 ModelOpt MIXED_PRECISION 格式(逐层标 NVFP4 或 FP8),SGLang 自动识别为 modelopt_mixed。张量共 2,191 个:语言模型 1,843、视觉 333、MTP 15。
• NVFP4 校准数据:取自本模型自己的 RLOO 训练数据中答对且答案非空的完整轨迹(提示 + 思考 + 最终答案),512 条、约 173.5 万 token,单条最多 4,096 token;官方 GPQA、MMLU、LCB 的题目都不进校准。详见 evaluation/NVFP4_CALIBRATION_METHOD.md。三档的视觉塔(333 个张量)与 MTP(15 个张量)均为 BF16。
• 结构:Qwen3_5ForConditionalGeneration,64 层(线性注意力与全注意力 3:1 交替)。

NVFP4 三档冒烟:SGLang 免补丁加载、贪心对拍、图像请求通过。MTP 与 DFlash2 的速度和接受长度以单独的性能测评为准(C1–C16 均已跑完,见上文)。

启动命令

需官方 SGLang ≥ 0.5.19;vLLM ≥ 0.28(仅 MTP)。vLLM 在张量并行 ≥ 2 时对 NVFP4 开 MTP 有已知问题(vLLM #52480),NVFP4 用 vLLM 开 MTP 时请用单卡。INT8 W8A8 与 INT4 W4A16 只能用 vLLM 加载(在 0.28 上实测),命令见本节末尾。

MTP 和 DFlash2 互斥,一次启动只能开其中一种。 下面以 ./NVFP4/W4A16 为例,用 W4A4 或 W4A4-W8A8 档时把 --model-path 换成 ./NVFP4/W4A4 或 ./NVFP4/W4A4-W8A8;DFlash2 草稿用各档目录下的 DFlash2-FP8(./NVFP4/W4A16/DFlash2-FP8、./NVFP4/W4A4/DFlash2-FP8、./NVFP4/W4A4-W8A8/DFlash2-FP8)。NVFP4 三档的量化格式由 SGLang 自动识别,不需要加 --quantization;NVFP4 只在 SGLang 上测过。并发与上下文参数与性能测试一致,可按需调整。

SGLang · NVFP4 W4A16 · MTP(包内 BF16)(W4A4、W4A4-W8A8 把 W4A16 换成对应目录名)

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A16 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4

vLLM · NVFP4 · MTP(每档一条命令,任选其一;单卡运行;本仓 NVFP4 包未在 vLLM 上实测)

# W4A16
vllm serve ./NVFP4/W4A16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# W4A4
vllm serve ./NVFP4/W4A4 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# W4A4-W8A8
vllm serve ./NVFP4/W4A4-W8A8 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

SGLang · NVFP4 W4A16 · DFlash2(W4A4 把两处 W4A16 换成 W4A4)

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A16 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path ./NVFP4/W4A16/DFlash2-FP8 \
  --speculative-draft-model-quantization fp8 \
  --speculative-num-draft-tokens 8

SGLang · NVFP4 W4A4-W8A8 · DFlash2

python -m sglang.launch_server \
  --model-path ./NVFP4/W4A4-W8A8 \
  --context-length 32768 \
  --max-running-requests 16 \
  --mamba-ssm-dtype bfloat16 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path ./NVFP4/W4A4-W8A8/DFlash2-FP8 \
  --speculative-draft-model-quantization fp8 \
  --speculative-num-draft-tokens 8

本卡 NVFP4 的三项测评分数是在 SGLang 上开 DFlash2 测出的(NInfer 与 INT8 W8A8 不开投机,INT4 W4A16 开 MTP,详见各节);投机解码不改变目标模型的输出分布。采样默认值见 generation_config.json(temperature 1.0、top_p 0.95、top_k 20)。

SGLang / vLLM 看图:NVFP4/W4A16/、NVFP4/W4A4/、NVFP4/W4A4-W8A8/ 都是完整多模态权重:config.json 为 Qwen3_5ForConditionalGeneration 且带 vision_config,目录里有 preprocessor_config.json 与 video_preprocessor_config.json,权重索引覆盖全部 333 个视觉塔张量(指向同目录的 vision-bf16.safetensors)。上面的 SGLang、vLLM 命令原样即可接收图片,不需要额外参数;图片按 OpenAI 格式放在 POST /v1/chat/completions 的 image_url 里(请求示例见下文「NInfer · 视觉」,model 换成服务的模型名)。INT8/W8A8/ 自 2026-10-09 起也能看图(Qwen3_5ForConditionalGeneration;BF16 视觉塔在 model_visual.safetensors,附 preprocessor 配置):直接用下面的 INT8 vLLM 命令即可。INT4/W4A16/ 同样能看图(BF16 视觉塔在 model_visual.safetensors),用下面的 INT4 vLLM 命令即可。

NVFP4 · NInfer(.ninfer,官方 NInfer)

以 W4A4 档为例;用 W4A4-W8A8 档时把路径换成 ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer,--model-id 换成 qwen3.8-27b-coder390-w4a4-w8a8。

MTP-DFlash2 包内已包含 Q8 MTP 和 DFlash2 草稿，NInfer 不需要外挂草稿模型（这一点与上面 SGLang 需要传 DFlash2 草稿路径不同）；启动时用 --spec mtp 或 --spec dflash2 二选一，不加 --spec 即不开投机。

图片输入：NVFP4-NInfer/ 下的两个 V3 MTP-DFlash2 .ninfer 包(W4A4 与 W4A4-W8A8)自 2026-10-09 起包含视觉部分(295,720,448 字节,由 BF16 视觉权重量化:patch embedding 为 Q6,attention q/k/v 与 MLP fc1 为 Q4,其余投影为 Q5,merger 为 Q8,norm 与 bias 保持 BF16)。文本、MTP 与 DFlash2 权重逐字节不变。在下面任一 NInfer 命令后加 --vision,图片按 OpenAI 格式放在 POST /v1/chat/completions 的 image_url 里。不加 --vision 时和以前一样只跑文本。V2 包(-NInferV2-NOT-FOR-V3)未改动,仍不含视觉。2026-10-09 GPU 实测(单张 RTX PRO 6000,NInfer-all 64492cb 加 --vision,看图问题):两个 MTP-DFlash2 包都回答正确。W4A4-MTP-DFlash2:MTP(3 个草稿 token)接受长度 2.89,decode 153 tok/s;DFlash2(7 个草稿 token)接受长度 3.53,decode 185 tok/s。W4A4-W8A8-MTP-DFlash2:MTP 3.19,164 tok/s;DFlash2 3.74,195 tok/s。

NInfer · 不开投机

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8

NInfer · MTP(草稿长度 4)

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft

NInfer · DFlash2(草稿长度 8)

ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft

NInfer · 视觉(图片输入),每个包一条命令

每条命令就是上面对应的 NInfer 命令加 --vision,不需要别的文件:视觉部分已在每个 V3 .ninfer 包内,同目录的 vision-bf16.safetensors 运行时用不到。

# W4A4 · MTP-DFlash2 包 · 不开投机
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --vision

# W4A4 · MTP-DFlash2 包 · MTP
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft \
  --vision

# W4A4 · MTP-DFlash2 包 · DFlash2
ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-W4A4-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft \
  --vision

# W4A4-W8A8 · MTP-DFlash2 包 · 不开投机
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --vision

# W4A4-W8A8 · MTP-DFlash2 包 · MTP
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft \
  --vision

# W4A4-W8A8 · MTP-DFlash2 包 · DFlash2
ninfer-serve ./NVFP4-NInfer/W4A4-W8A8/Qwen3.8-27B-Coder390-EfficientThink-W4A4-W8A8-MTP-DFlash2.ninfer \
  --host 127.0.0.1 --port 19931 --device 0 \
  --model-id qwen3.8-27b-coder390-w4a4-w8a8 \
  --max-context 102400 --kv-capacity 819200 --max-concurrency 8 \
  --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 8 --lm-head-draft \
  --vision

图片请求:按 OpenAI 格式在 POST /v1/chat/completions 里放 image_url;url 填图片的 http(s):// 链接,也可以传本地图片的数据 URI;model 要与 --model-id 一致。

curl http://127.0.0.1:19931/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b-coder390-w4a4",
    "messages": [{"role": "user", "content": [
      {"type": "image_url", "image_url": {"url": "https://example.com/demo.png"}},
      {"type": "text", "text": "Describe this image."}
    ]}],
    "max_tokens": 2048
  }'

健康检查:curl -sf http://127.0.0.1:19931/health。对话接口为 POST /v1/chat/completions,请求里的 model 要与 --model-id 一致。

vLLM · INT8 W8A8(不开投机)

VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel \
VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT8/W8A8 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int8 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3

RTX PRO 6000 Blackwell 上 vLLM 的 Cutlass INT8 内核不可用,不关掉就起不来,须用 VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel 关掉;关掉后改走 Triton INT8 内核(启动日志里能看到 TritonInt8ScaledMMLinearKernel)。VLLM_USE_FLASHINFER_SAMPLER=0 与测试时一致。在 vLLM 0.28 上实测;SGLang 不能加载本包(ModelOptFp8Config 报错)。--reasoning-parser qwen3 会把思考内容单独放进 reasoning_content(测分时未开)。

vLLM · INT8 W8A8 · MTP(外挂草稿 INT8/W8A8/MTP-BF16/)

VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel \
VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT8/W8A8 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int8 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","model":"./INT8/W8A8/MTP-BF16","num_speculative_tokens":3}'

在 RTX PRO 6000 Blackwell 上按上面的命令设置 VLLM_DISABLED_KERNELS=CutlassInt8ScaledMMLinearKernel 和 VLLM_USE_FLASHINFER_SAMPLER=0。新表的两行都用了这两项设置。

vLLM · INT4 W4A16 · MTP(内置头)· 文本与图像

VLLM_USE_FLASHINFER_SAMPLER=0 \
vllm serve ./INT4/W4A16 \
  --trust-remote-code --dtype bfloat16 \
  --max-model-len 102400 --served-model-name int4 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

图片按 OpenAI 格式放在 image_url 里,用同一个服务即可,不需要额外参数:

curl -s http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "int4",
  "messages": [{"role": "user", "content": [
    {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
    {"type": "text", "text": "描述这张图片。"}
  ]}],
  "max_tokens": 2048
}'

去掉 --speculative-config 即不开投机。在 vLLM 0.28 上实测(VLLM_USE_FLASHINFER_SAMPLER=0 与实测设置一致);SGLang 与 NInfer 未用此包测试。--reasoning-parser qwen3 会把思考单独返回。

其他文件纵览

NVFP4/W4A16/(主包 17 个文件,20,613,406,953 字节;DFlash2 子目录另列在下面)

NVFP4/W4A16/SHA256SUMS 覆盖下表前 16 个文件,其自身 SHA256 为 ecc28d31be0d97e23509af61625251af14980db5b67189a1d1ca75a3fc9652f3。mtp-bf16.safetensors 与主仓各档中的同名文件逐字节相同。

NVFP4/W4A16/DFlash2-FP8/(4 个文件,2,407,384,680 字节)

与主仓 FP8/DFlash2-FP8/ 逐字节相同。

NVFP4/W4A4/(主包 17 个文件,20,613,386,462 字节;DFlash2 子目录另列在下面)

NVFP4/W4A4/SHA256SUMS 覆盖下表前 16 个文件,其自身 SHA256 为 5aea8bb9c1f1979661bfa5c1216de7aafeb64cd708aa8cc71f3b185c224cfde0。mtp-bf16.safetensors 与主仓各档中的同名文件逐字节相同。

NVFP4/W4A4/DFlash2-FP8/(4 个文件,2,407,384,680 字节)

与主仓 FP8/DFlash2-FP8/ 逐字节相同。

NVFP4/W4A4-W8A8/(主包 18 个文件,23,769,634,648 字节;DFlash2 子目录另列在下面)

NVFP4/W4A4-W8A8/SHA256SUMS 覆盖下表前 17 个文件,其自身 SHA256 为 c177f4907a7dd037efe91b8d34927dd3a3624227f259238b82db0f12f8b0b8e2。mtp-bf16.safetensors 与主仓各档中的同名文件逐字节相同。

NVFP4/W4A4-W8A8/DFlash2-FP8/(4 个文件,2,407,384,680 字节)

与主仓 FP8/DFlash2-FP8/ 逐字节相同。

evaluation/

NVFP4 W4A16、W4A4 与 W4A4-W8A8 三档的成绩、性能与校准记录,evaluation/SHA256SUMS 为该目录的校验清单。NVFP4_*.results.summary.json 为 NVFP4 三档三项测评的原始汇总(NVFP4_MIXED_* 为 W4A4-W8A8 档)。

校验

(cd NVFP4/W4A16 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4/W4A4 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4/W4A4-W8A8 && sha256sum -c SHA256SUMS && cd DFlash2-FP8 && sha256sum -c SHA256SUMS)
(cd NVFP4-NInfer/W4A4 && sha256sum -c SHA256SUMS)
(cd NVFP4-NInfer/W4A4-W8A8 && sha256sum -c SHA256SUMS)
(cd INT8/W8A8 && sha256sum -c SHA256SUMS && cd MTP-BF16 && sha256sum -c SHA256SUMS)
(cd INT4/W4A16 && sha256sum -c SHA256SUMS)
(cd evaluation && sha256sum -c SHA256SUMS)

NInfer 加载说明

• 只有 NInfer 引擎能加载 .ninfer;SGLang、vLLM、llama.cpp、transformers 都不能读。
• 使用官方 NInfer(Neroued/ninfer 的 master 分支)的 ninfer-serve,无需打补丁(只有由 GGUF 转换的 .ninfer 需用 NInfer-all)。实测硬件为单张 RTX PRO 6000 Blackwell(算力 12.0);NInfer 并发上限为 8。
• NVFP4-NInfer/W4A4/ 与 NVFP4-NInfer/W4A4-W8A8/ 各有一个带 MTP 与 DFlash2 的 .ninfer 文件:文本路径与 2026-10-05 测分时相同,包内另含 Q8 MTP、DFlash2 草稿、proposal head,以及 2026-10-09 加入的视觉部分,加 --vision 可输入图片;同一文件夹的 vision-bf16.safetensors 是这部分的 BF16 原始权重,运行时不需要。MTP 投影为 Q8:其中几个矩阵形状在 NInfer 里没有 BF16 内核,用 BF16 会在建计算图时退出。
• MTP 与 DFlash2 一次启动只能选一种(--spec mtp 或 --spec dflash2);--lm-head-draft 须与 --spec 一起用。MTP 草稿长度可取 1–5,DFlash2 可取 1–15;上文「启动命令」中的 NInfer 命令用测速时的 4 与 8。DFlash2 草稿已打在包里,不需要另载 DFlash2-FP8/ 目录。
• 下面的上下文参数已验证能启动:--max-context 102400 --kv-capacity 819200 --max-concurrency 8 --kv-dtype fp8。--kv-capacity 须落在 max-context 到 max-context × max-concurrency 之间,819,200 正好是 8 × 102,400。
• NVFP4 目录下的 HF 包仍用 SGLang(可开 MTP 或 DFlash2);与 NInfer 包互不通用。

NInfer V2 MTP 包(支持看图,不含 DFlash2,Blackwell)

两种 V2 包,别混用(都不能在现在的 V3 引擎上运行,即当前官方 NInfer 与 NInfer-all):

V100 包的用法见下一节,保持不变。

2026-10-11 新增两个 NInfer V2 容器包,用于想要更小体积的 Blackwell 显卡(如 32GB 卡):包内只有 NVFP4 文本权重、Q8 MTP 头和视觉部分,不含 DFlash2 草稿。文件比 V3 MTP-DFlash2 包小约 2.3 GB。

• 需要的引擎:官方 NInfer 在切换到 V3 之前的版本,commit 99f08f00(2026-09-13),只能在 Blackwell(sm_120)显卡上编译运行,需要 CUDA 13.1 或更新。现在的官方 NInfer、NInfer-all 只读 V3,会拒绝 V2 包;V100 的 ninfer-v100 分支也不适用这两个包。
• 功能:支持 --spec mtp 和 --vision;不支持 --spec dflash2。
• 两档的区别:W4A4 与 W4A4-W8A8 的源量化不同,但转换成 NInfer 包后,包内格式完全一样(张量数量、格式、大小逐项一致,只是数值不同)。
• 包内聊天模板与本仓 BF16/NVFP4 的 chat_template.jinja 相同(该版本 NInfer 只认这份模板)。

编译(Blackwell,CUDA 13.1+):

git clone https://github.com/Neroued/ninfer.git && cd ninfer && git checkout 99f08f00
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=120a
cmake --build build --target ninfer-serve

实测启动命令(MTP + 看图):

./build/apps/ninfer-serve ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-EfficientThink-NVFP4-W4A4-MTP-NInferV2.ninfer \
  --host 127.0.0.1 --port 8000 \
  --max-context 102400 --kv-capacity 102400 --max-concurrency 1 --kv-dtype fp8 \
  --spec mtp --draft-tokens 4 --lm-head-draft --vision

不需要看图时去掉 --vision,显存约少 0.9 GiB。

实测(2026-10-11,单张 RTX PRO 6000 Blackwell,上述命令,C1 依次发 4 个编程请求,生成上限 4,096,开思考,temperature 1.0、top_p 0.95、top_k 20):

同条件不开投机时 W4A4 为 69 tok/s。

显存(W4A4,nvidia-smi 读数,含 CUDA 上下文;FP8 KV,单并发,--kv-capacity 等于上下文长度,KV 在启动时一次分配;每档都实际跑过约 30K / 61K / 97K token 的长提示词):

以上显存数字为 --draft-tokens 3 时测得;--draft-tokens 4 --lm-head-draft、100K 上下文时为 25,641 MiB。另外服务默认会锁定约 9 GiB 主机内存(主机 KV 缓存 8 GiB 等)。

NInfer V2 包(仅 Tesla V100)

文件名带 NInferV2-NOT-FOR-V3 的四个包只给 Tesla V100 用（ninfer-v100 分支）；其他显卡请用不带这个后缀的包。

• 文件名含 NInferV2-NOT-FOR-V3 的 4 个包是 NInfer V2 容器,只给 Tesla V100 的 ninfer-v100 分支用。官方 NInfer(ninfer-serve)只读 V3,会拒绝这些文件;其他显卡请用上面不带 NInferV2 的 V3 包。不要把 V2 包改名成 V3 文件名,也不要用它们替换 V3 包。
• 两个纯文本包(Qwen3.8-27B-Coder390-W4A4-NInferV2-NOT-FOR-V3.ninfer、Qwen3.8-27B-Coder390-W4A4-W8A8-NInferV2-NOT-FOR-V3.ninfer)不含 MTP 与 DFlash2,不要加 --spec;两个 MTP-DFlash2 包(…-MTP-DFlash2-NInferV2-NOT-FOR-V3.ninfer)可加 --spec mtp 或 --spec dflash2,包内已包含 MTP 和 DFlash2 草稿,不需要外挂草稿模型。四个 V2 包都不含视觉部分,不能输入图片(不支持 --vision);需要看图请用 V3 包。
• V100 上没有 NVFP4 KV 与 K8V4 KV,不要用 V3 的 --kv-dtype rk8v4 或 --kv-capacity 819200;在 V100 上起服务时加 --prefill-chunk 2048。
• V2 包未单独测评,只做了 SHA256 校验。SHA256SUMS 里 V2 的两行放在 # 注释行之后,与 V3 分开。

# V100 ninfer-v100 分支 · 纯文本包(不加 --spec)
ninfer ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-W4A4-NInferV2-NOT-FOR-V3.ninfer \
  --prompt "用三句话说明 prefill 和 decode。" \
  --max-context 8192 \
  --max-new 256

# V100 ninfer-v100 分支 · MTP-DFlash2 包(开 MTP)
ninfer ./NVFP4-NInfer/W4A4/Qwen3.8-27B-Coder390-W4A4-MTP-DFlash2-NInferV2-NOT-FOR-V3.ninfer \
  --prompt "用三句话说明 prefill 和 decode。" \
  --max-context 8192 \
  --max-new 256 \
  --spec mtp --draft-tokens 3

NVFP4-NInfer/W4A4/(5 个文件,68,925,700,831 字节)

NVFP4-NInfer/W4A4-W8A8/(5 个文件,68,925,700,846 字节)

INT8 W8A8 说明

• 文本 + 图像(2026-10-09 起):架构为 Qwen3_5ForConditionalGeneration(文本 64 层,线性注意力与全注意力 3:1 交替)。视觉塔取自 BF16 档,以 BF16 单独存为 model_visual.safetensors(333 个张量,921,497,224 字节),并附 preprocessor_config.json / video_preprocessor_config.json;视觉塔不量化。8 个 INT8 权重文件未改动。包内不含 MTP 头(用 MTP-BF16/),也不含 DFlash2 草稿。2026-10-09 在一台 RTX PRO 6000 上实测(vLLM 0.28,关掉 Cutlass INT8 内核,MTP-BF16/、3 个草稿 token):看图问题回答正确,文本输出正常,接受长度 3.21,单请求 decode 86 tok/s。
• 量化:SmoothQuant + INT8 W8A8。权重按输出通道量化为 INT8,激活在推理时按 token 动态量化为 INT8;SmoothQuant 的预缩放已折进权重,推理框架不需要补丁。量化元数据为 compressed-tensors(int-quantized)格式。
• 层分布:MLP gate/up/down、全注意力 q/k/v/o、线性注意力 in_proj_qkv / in_proj_z / out_proj 共 400 个 Linear 为 INT8;线性注意力 in_proj_a / in_proj_b 与 lm_head 共 97 个保持 BF16;线性注意力的 A_log、dt_bias、norm、conv1d 等控制分支也保持 BF16。
• 校准与误差:512 条校准样本;留出集 NLL 由 BF16 的 0.50535 升到 0.52812,困惑度之比 1.023(在把预缩放折进权重之前测得)。
• 加载:只在 vLLM 0.28 上验证,不适用于 SGLang 与 NInfer;VLLM_LOAD.json 记录了本包的加载条件。RTX PRO 6000 Blackwell 上须关掉 Cutlass INT8 内核,见上文「启动命令」。
• MTP 草稿:INT8/W8A8/MTP-BF16/ 只有 BF16 的 MTP 张量。2026-10-09 在一台 RTX PRO 6000 上用 vLLM 0.28 实测(关掉 Cutlass INT8 内核,num_speculative_tokens=3):平均接受长度 2.74–2.90。

INT8/W8A8/(19 个文件,30,422,466,299 字节)

INT8/W8A8/MTP-BF16/(3 个文件,849,403,270 字节)

INT8/W8A8/MTP-BF16/SHA256SUMS 覆盖下表前 2 个文件,其自身 SHA256 为 313b058349deba83f05adb44b79f2f76ae668337834086f8dbcfacba131bf403。

INT4 W4A16 说明

• 文本与图像:架构为 Qwen3_5ForConditionalGeneration(64 层文本,线性注意力与全注意力 3:1 交替)。BF16 视觉塔单独存为 model_visual.safetensors(333 个张量,921,497,224 字节,与 INT8/W8A8/ 中的是同一个文件),附 preprocessor_config.json / video_preprocessor_config.json。内置 MTP 头为 BF16(model_mtp.safetensors,15 个张量)。不含 DFlash2 草稿。
• 量化:用 llm-compressor 做 GPTQ W4A16,INT4 权重,对称,分组大小 128,静态激活顺序,dampening 0.01;激活保持 BF16。lm_head、词嵌入、MTP 头与视觉塔不量化。校准数据:用本模型 RL 训练 rollout 重新整理的 512 条 × 4,096 token。量化元数据为 compressed-tensors(pack-quantized)格式;recipe.yaml 是 llm-compressor 配方。
• BPW 5.25:model.safetensors(17,646,891,544 字节,只含文本权重)× 8 ÷ 文本参数数 26,895,998,464;MTP 与视觉文件不计入。
• 看图 + MTP 检查(2026-10-09,单张 RTX PRO 6000,vLLM 0.28,3 个草稿 token):开关思考两种情况下看图问题都回答正确。
• 加载:只在 vLLM 0.28 上验证过。VLLM_LOAD.json 记录了加载设置。

INT4/W4A16/(14 个文件,19,437,985,972 字节)

已知局限

• 长尾仍未消除:思考 ≥48K 的题正确率明显偏低(W4A16 的 LCB ≥48K 为 4/10,W4A4 为 7/11,W4A4-W8A8 为 4/9),LCB 仍有 4 道(W4A16)、3 道(W4A4)与 2 道(W4A4-W8A8)写满 94K。
• 成绩依赖上述口径(100K 上下文、94,208 生成上限、不计超时),不宜与其他口径直接比较;MMLU 有采样抖动。
• NVFP4 三档的性能在单张 RTX PRO 6000 Blackwell + SGLang 上测得,三档各用自己的包;NVFP4 未在 vLLM 上测试;DFlash2 只在 SGLang 上验证。NInfer 两档在 NInfer 上测,INT8 W8A8 与 INT4 W4A16 在 vLLM 0.28 上测。
• NVFP4 三档的 GPQA(170、167、172)低于主仓 BF16 / FP8 的 178,MMLU(439、440、442)略低于 445–450,LCB(91、92、88)在 90 上下;校准数据来自本模型自己的 RLOO 训练数据,不含官方测评题。NVFP4 只在 SGLang 上做过免补丁加载、贪心对拍和图像请求冒烟,MTP 与 DFlash2 以完整的性能测评为准。
• INT8 W8A8 只能用 vLLM(分数在文本上测得;2026-10-09 加入了图像输入);GPQA 175 略低于原版 Qwen3.8 27B(FP8)的 177;速度为快测(每档请求数等于并发数,生成上限 512),不能与其他档直接比较;GPQA / MMLU 思考长度按原文用 BF16 tokenizer 重新计数,LCB 没有保存思考原文,没有思考长度分位数。
• 输出请自行核实,尤其是高风险场景。

许可证与致谢

本模型采用 Apache-2.0 许可证,遵循基座 Qwen3.8-27B 的许可。

感谢 Qwen 团队提供基座模型与官方 FP8 方案;感谢 Opus5.5、GPT6Astra、Grok4.7、DSV4Pro、K3 在出题、金标、教师轨迹、数据核对与 RLOO 价值评审中的工作;感谢 SGLang、vLLM 与 DFlash 社区。