Qwen3.8-27B — NVFP4 + MTP + VL (GGUF, 1M YaRN)

Community llama.cpp / Ollama pack of Qwen/Qwen3.8-27B.

This is not the official Transformers / vLLM / SGLang checkpoint. The official repo is BF16 safetensors (~55 GB) for Hugging Face Transformers and dedicated servers. This repo is a quantized GGUF split: NVFP4 language + trained MTP, plus an F16 CLIP projector, with GGUF metadata patched so runners that honor n_ctx_train will accept 1,048,576 tokens (official static YaRN).

License is the same Apache License 2.0 as Qwen/Qwen3.8-27B. The LICENSE file in this repo is a verbatim copy of that official text (including the Alibaba Cloud 2026 appendix notice).

What you get

The projector is not inside the 27B GGUF. That file is qwen35 language + MTP. CLIP lives in its own mmproj GGUF (general.type = mmproj, F16). You do not convert the official tower, merge tensors, or use ADAPTER (that is LoRA). Point the runner at the file we already shipped: --mmproj or a second FROM.

mmproj/ is only a folder so Hugging Face indexes the 27B language file (Ollama app filter, ~27B / qwen35) instead of the 0.5B CLIP file. Same projector weights; not a second copy you have to “embed.”

NVFP4 applies only to language (and MTP) matmuls. The vision tower stays F16. 1M context is not extra trained weights; it is rope/context metadata so the runner will allocate a 1M window.

Why this pack exists

Qwen3.8-27B is a native VLM (images and video) with thinking on by default, trained MTP, 262,144 native context, and an official path to 1,000,000 tokens via static YaRN (factor: 4.0, original_max_position_embeddings: 262144). Qwen’s hosted API is expected to default to 1M; the open Transformers card still ships 262K in config.json and documents YaRN as an override.

Goal for local serve (DGX Spark / Ollama 0.32 / llama.cpp):

1. Keep NVFP4 so ~27B fits in ~16 GB instead of ~55 GB BF16.
2. Keep MTP so speculative decode works (--spec-type draft-mtp).
3. Keep vision (official CLIP projector).
4. Make 1M actually loadable in Ollama, which caps num_ctx at GGUF n_ctx_train.

Strategy (what we did and did not do)

What we did not do

• We did not retrain Qwen3.8.
• We did not NVFP4-quantize from BF16 on this tree. Current llama.cpp llama-quantize has no NVFP4 target in QUANT_OPTIONS (NVFP4 exists as a runtime ftype). Do not run llama-quantize … NVFP4 and expect a good file.
• We did not bake 1M into the tensors. KV at 1M is still ~61–64 GiB FP16 plus the ~16 GB weights.
• We did not put vision into the NVFP4 blob. That would be the wrong precision and the wrong packaging.

Pipeline

Qwen/Qwen3.8-27B (official BF16 VLM + MTP, 262K native / 1M YaRN recipe)
        │
        ├─ language+MTP NVFP4 GGUF
        │     Avifenesh/Qwen3.8-27B-NVFP4-MTP-GGUF
        │     (text-only file; VL tower not inside)
        │
        ├─ CLIP mmproj F16
        │     convert official HF with llama.cpp / Ollama
        │     (language_model_only: false, vision_config present)
        │
        └─ copy language GGUF → rewrite KV only
              qwen35.context_length = 1048576
              rope.scaling.type = yarn
              rope.scaling.factor = 4.0
              rope.scaling.original_context_length = 262144
              freq_base 1e7, mrope sections [11, 11, 10] already present

Ollama is assembled with two FROM lines (language blob + projector blob). ADAPTER is LoRA and is the wrong instruction. RENDERER/PARSER qwen3-vl is for the older qwen3-vl arch, not qwen35.

Why the YaRN rewrite

On Ollama 0.32, PARAMETER num_ctx 1048576 against the stock 262K NVFP4 GGUF is rejected:

requested context size too large for model num_ctx=1048576 n_ctx_train=262144

Unified memory was not the limiter. After the KV patch, ollama show reports context length 1048576. The shipped Modelfile still defaults to num_ctx 262144 so a casual chat does not allocate a 1M KV cache. Pass options.num_ctx: 1048576 only when you want that window.

Qwen’s own card: static YaRN can hurt short prompts. Prefer 8K–262K for chat; use 1M for long jobs.

Use

Needs a llama.cpp / Ollama build that understands qwen35, NVFP4, mmproj, and draft-mtp.

llama.cpp

huggingface-cli download mixbits/Qwen3.8-27B-NVFP4-MTP-VL-GGUF --local-dir ./qwen38-nvfp4

llama-server \
  -m ./qwen38-nvfp4/Qwen3.8-27B-NVFP4-Q5K-mtp-yarn1m.gguf \
  --mmproj ./qwen38-nvfp4/mmproj/mmproj-Qwen3.8-27B-F16.gguf \
  --spec-type draft-mtp --spec-draft-n-max 4 \
  -c 8192

Use -c 262144 or -c 1048576 only if you have the RAM.

Ollama

# after download, from the folder that contains both GGUFs + Modelfile
ollama create qwen38-27b-nvfp4-mtp-vl -f Modelfile

FROM ./Qwen3.8-27B-NVFP4-Q5K-mtp-yarn1m.gguf
FROM ./mmproj/mmproj-Qwen3.8-27B-F16.gguf

PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER min_p 0.0
PARAMETER repeat_penalty 1.0
PARAMETER presence_penalty 0.0
PARAMETER num_ctx 262144
PARAMETER draft_num_predict 4

draft_num_predict 4 → --spec-type draft-mtp on Ollama 0.32.x.

Thinking is on by default (same as the official model). A vision request with thinking enabled can return HTTP 200 and empty content. Use instruct / "think": false for images.

Open WebUI: leaving num_ctx as Default uses the Modelfile (262K). Toggling the control pre-fills 2048 and will silently shrink the window. Set num_ctx to 1048576 on this model only if you intend a ~64 GiB KV alloc. Do not raise a global Ollama OLLAMA_CONTEXT_LENGTH to 1M.

Sampling (from the official card)

Official reasoning_effort (xhigh / medium / low) and preserve_thinking are Transformers / Qwen Cloud features. GGUF+Ollama exposes thinking on/off more coarsely.

Validation (DGX Spark, GB10, Ollama 0.32.11)

Official text and VL benchmark tables are for the BF16 model, not this quant.

Official model (condensed)

From Qwen/Qwen3.8-27B:

• Causal LM + vision encoder; 27B; hidden 5120; 64 layers; vocab 248,320 (padded)
• Hybrid layout: gated DeltaNet + gated attention (RoPE dim 64, partial_rotary_factor 0.25)
• MTP trained with multiple steps
• Native VLM (images and video); thinking on by default
• Context: 262,144 native, extensible to 1,000,000 with YaRN

YaRN fields used here match the official recipe:

{
  "mrope_interleaved": true,
  "mrope_section": [11, 11, 10],
  "rope_type": "yarn",
  "rope_theta": 10000000,
  "partial_rotary_factor": 0.25,
  "factor": 4.0,
  "original_max_position_embeddings": 262144
}

Video longest_edge tuning from the official card is a Transformers / vLLM concern; this GGUF pack is validated for images, not hour-scale video.

Attribution

• Qwen/Qwen3.8-27B — architecture, training, license, YaRN recipe, sampling
• Avifenesh/Qwen3.8-27B-NVFP4-MTP-GGUF — NVFP4 + MTP language GGUF this pack copies and metadata-patches
• This repo (mixbits) — YaRN/1M GGUF KV, official CLIP mmproj attach, Ollama dual-FROM layout, Spark validation

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}