Qwen3.8-27B-MLX-bf16

<!-- base_model_relation is set explicitly: with no quantization key in config.json the Hub would default this repo into the base model's "Finetunes" tree, which is wrong (mlx-community/Qwen3.6-27B-bf16 sits there today). "quantized" is the Hub's bucket for inference-format conversions — GGUF F16 repos live there too. UPLOAD NOTE: mlx_lm.convert writes its own stub README.md into the output directory — replace it with this card when uploading. -->

Qwen3.8-27B — MLX, bf16

This model was converted to MLX format from Qwen/Qwen3.8-27B using mlx-lm. Refer to the original model card for more details on the model.

The conversion is lossless for the language model (~54 GB on disk): every weight is bit-identical to the source, except the RMSNorm gains, which are stored in the mathematically equivalent 1 + w form that the reference computes at runtime (verified exact: embeddings, lm_head, linear and A_log tensors bit-identical; every checked norm exactly source + 1.0). No quantization.

This conversion is text-only. The base model is a native vision-language model; mlx-lm's qwen3_5 implementation strips the vision encoder (vision_tower.*) and the multi-token-prediction drafter (mtp.*) at load time, so this artifact does not accept images or videos and does not do speculative decoding. For vision use, run the original checkpoint with Transformers, vLLM, or SGLang. (The pipeline_tag here is text-generation for that reason, while the base card is image-text-to-text.)

Model description (inherited from the base card)

Qwen3.8-27B is the compact dense model of the Qwen3.8 generation, built on the Qwen3.5 architecture:

• 27B parameters, 64 layers: 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)) — a 3:1 hybrid of linear attention and full attention
• Gated DeltaNet: 48 V-heads / 16 QK-heads, head dim 128
• Gated Attention: 24 Q-heads / 4 KV-heads, head dim 256, RoPE on 64 dims
• Vocab 248,320; context 262,144 native (extensible to 1M — extended-context serving beyond 256K is not configured in this conversion)
• Thinking mode on by default, reasoning_effort control (xhigh default / medium / low), preserve_thinking on by default

See the base model card for benchmark results. Text benchmarks apply to this conversion in principle (identical bf16 numerics); vision benchmarks do not (no vision tower).

Use with mlx-lm

pip install -U mlx-lm

mlx_lm.generate --model malekoo/Qwen3.8-27B-MLX-bf16 --prompt "Explain KV caches briefly." --max-tokens 2048

from mlx_lm import load, generate

model, tokenizer = load("malekoo/Qwen3.8-27B-MLX-bf16")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain KV caches briefly."}],
    add_generation_prompt=True,
)
text = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=True)

OpenAI-compatible server:

mlx_lm.server --model malekoo/Qwen3.8-27B-MLX-bf16 --port 8080

In requests, set "model" to the id reported by /v1/models (the model path) — mlx-lm treats an unrecognized name as a Hub repo to download.

Thinking control goes through chat_template_kwargs, as on the base model: enable_thinking (default true), preserve_thinking (default true), and reasoning_effort (xhigh default, medium, low) — all verified effective against this repo's chat template. The generation prompt ends inside an open <think> block (thinking mode), so reasoning-channel splitting requires a parser aware of the pre-opened tag.

Recommended sampling (from the base card)

• Thinking mode: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (non-thinking) mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Qwen notes the model operates in thinking mode by default, emitting <think>...</think> before the answer; mlx-lm's server surfaces this as the reasoning channel.

Conversion provenance

• Source: Qwen/Qwen3.8-27B (bf16 safetensors, released 2026-08-14)
• Converted with mlx_lm.convert (no -q); the vision encoder (333 tensors) and MTP drafter (15 tensors) are dropped, the remaining 851 text tensors verified as described above
• Tokenizer re-serialized by transformers 5; verified id-identical to the source tokenizer on text, code, CJK, emoji, and chat-template inputs
• Toolchain: mlx 0.32.1.dev20260814+3d23f7d87 (built from ml-explore/mlx main), mlx-lm at ml-explore/mlx-lm main 254d153 (conversion code path verified identical to upstream main)
• Hardware: Apple silicon (macOS), Metal backend
• Measured on this build: wikitext-2 (test) perplexity 6.9352 — disjoint 2048-token windows, no overlap, single token stream. Decode ~10 tok/s on an M5 Max MacBook Pro at ~54 GB peak memory. A 4-bit quantization measured in the identical harness is at malekoo/Qwen3.8-27B-MLX-4bit (ppl 7.0871, ~15 GB, ~33 tok/s).

License and attribution

Apache-2.0, inherited from the base model. Copyright the Qwen team; this repository is a format conversion and claims no additional rights.

Citation

@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}
}