Huihui-Qwen3.8-27B-abliterated-FP8

Block-wise FP8 (e4m3) quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated, an abliterated version of Qwen/Qwen3.8-27B.

52 GiB (BF16) → 29 GiB (FP8). The vision tower, all normalization layers, and the MTP head's own norms/fusion layer stay in BF16, so vLLM's multimodal path and MTP speculative decoding both keep working.

Update — 2026-09-08

Re-quantized from the updated base model (commit 739e3c5, uploaded 2026-08-24), in which huihui-ai narrowed the abliteration to a band of middle layers instead of the whole upper stack. The narrower band keeps more of the parent model's behaviour intact while still removing refusals.

Measured against the parent FP8 checkpoint, this revision modifies only the residual-writing projections (mlp.down_proj, self_attn.o_proj / linear_attn.out_proj):

(huihui's model card describes the band as "layers 18 to 51"; the per-layer weight diff puts the lower edge one layer earlier, at 17.) For layers 17–51 the ablation transform is byte-identical to the previous revision — only the range changed, releasing layers 15–16 and 52–63 back to parent values.

The quantization recipe, module set and on-disk layout are unchanged from the previous revision; only the underlying weights are newer. The earlier revision of this repo is still reachable in the git history if you want the wider ablation.

Quantization details

The layout is identical to the official FP8 release of this model (Qwen/Qwen3.8-27B-FP8):

Unlike the earlier Qwen3.6 FP8 releases, the official Qwen3.8 FP8 checkpoint also quantizes part of the MTP head (mtp.layers.0.self_attn.* and mtp.layers.0.mlp.*) — this release matches that. The quantized-module set is taken verbatim from the reference checkpoint rather than chosen by hand, so it tracks whatever the official release does.

Deliberately left in BF16 — quantizing these breaks the model:

• Gated DeltaNet internals: conv1d, in_proj_a, in_proj_b, A_log, dt_bias, norm
• All input_layernorm / post_attention_layernorm / q_norm / k_norm
• lm_head, embed_tokens
• The entire vision tower (model.visual.*)
• The MTP head's own norms and mtp.fc (fusion layer)

Verification

The quantizer was validated against Qwen/Qwen3.8-27B-FP8 (the FP8 release of this model's unablated parent). Every module abliteration did not touch quantizes to the reference FP8 bit-for-bit; a full per-layer sweep of the residual-writing projections confirms the 17–51 band precisely:

layers 0–16, 52–63 + entire MTP head   — bit-identical to reference
layers 17–51 (down_proj / out-proj)    — differ, as expected
gate_proj / up_proj / q_proj (any layer) — bit-identical to reference

Tensor names, dtypes and shapes all match the reference (1199 BF16 + 407 FP8 = 1606 tensors; quantization_config identical to both the reference and the previous revision).

Usage (vLLM)

vllm serve munekazu/Huihui-Qwen3.8-27B-abliterated-FP8 \
  --quantization fp8 \
  --dtype bfloat16 \
  --tensor-parallel-size 2 \
  --max-model-len 262144 \
  --kv-cache-dtype fp8 \
  --trust-remote-code \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Tested on vLLM v0.24.0.

Chat template / thinking

Uses the model's bundled chat_template.jinja, which is reasoning_effort-based (low / medium / high / xhigh) rather than the older enable_thinking boolean. Pass e.g. "chat_template_kwargs": {"reasoning_effort": "low"} in requests, or set --default-chat-template-kwargs server-side.

⚠️ Known issue observed on the previous revision (vLLM v0.24.0): with --reasoning-parser qwen3, reasoning_content comes back empty in both streaming and non-streaming responses, even though the model is actually thinking (verified via the raw /v1/completions endpoint, which shows a normal <reasoning>...</think> generation). Feeding the exact same generated text to the parser directly (outside the server) extracts reasoning/content correctly, so the bug looks like it's in how vLLM v0.24.0's decode loop drives the parser during generation, not in the parser's extraction logic itself. Not root-caused further. Answer quality is unaffected either way; you just don't get the reasoning trace back over the API.

Hardware notes

Verified on 2× RTX 3090 (Ampere, sm_86) with TP=2, ~22 GB per card at --gpu-memory-utilization 0.92 and 262K context with fp8 KV cache.

Ampere has no FP8 tensor cores, so vLLM falls back to the Marlin W8A16 kernel (weight-only decompression). You get the memory saving but no FP8 compute speedup — expect this warning, which is normal:

Your GPU does not have native support for FP8 computation but FP8 quantization
is being used. Weight-only FP8 compression will be used leveraging the Marlin kernel.

On Ada / Hopper / Blackwell (sm_89+) FP8 runs natively.

MTP speculative decoding works: measured mean acceptance length 2.22 on 2× 3090.

Inherited warnings from the base model

This is a quantization of an abliterated (uncensored) model. The safety filtering of the base model has been significantly reduced. The warnings from huihui-ai apply unchanged:

• Risk of sensitive or controversial outputs — review generated content carefully
• Not suitable for all audiences — may be inappropriate for public settings or underage users
• Legal and ethical responsibility rests with the user — ensure compliance with local law
• Intended for research, testing and controlled environments, not unsupervised production use
• No default safety guarantees — this model has not undergone safety optimization

Quantization does not change the behaviour of the base model in this respect. Credit for the base model and its abliteration goes to huihui-ai; please support their work there.