Qwen3.8-27B-MLX-8bit

Qwen3.8-27B MLX 8-bit

Unofficial community conversion of Qwen/Qwen3.8-27B for Apple Silicon using MLX and MLX-VLM.

This is a format conversion and affine 8-bit quantization of the language model weights. The vision tower remains in BF16. No fine-tuning, merging, abliteration, distillation, or additional training was performed.

This repository is not an official Qwen, Alibaba Cloud, Apple, MLX, MLX-VLM, Hugging Face, or LM Studio release, and those parties do not endorse it.

Model details

The upstream model is a 27B native vision-language model. It supports text, image, and video understanding, flexible thinking control, and agentic/tool use through compatible serving stacks. See the Qwen3.8-27B model card for the architecture, upstream evaluations, recommended sampling settings, intended uses, and model-family limitations.

What changed

• The language-model tensors were converted to MLX and quantized with affine 8-bit weights using a group size of 64.
• The vision encoder was retained in BF16 by the converter.
• tokenizer_config.json embeds the same chat template also distributed as chat_template.jinja, improving compatibility with offline and local serving stacks.
• The generated configuration records the quantization parameters.
• Release documentation and SHA-256 checksums were added.

The exact file-level changes and reproducible conversion command are recorded in MODIFICATIONS.md.

Requirements

• A Mac with Apple Silicon.
• A current version of mlx-vlm; this conversion was produced and validated with mlx-vlm==0.6.15 and mlx==0.32.1.
• Sufficient unified memory. Short-context validation peaked at approximately 35.8 GB. At least 48 GB of unified memory is recommended for practical use. Long contexts and concurrent requests require substantially more memory.

Use with MLX-VLM

Install the runtime:

python -m pip install -U "mlx-vlm>=0.6.15"

Text

python -m mlx_vlm.generate \
  --model z4bfg/Qwen3.8-27B-MLX-8bit \
  --prompt "Explain affine weight quantization in three concise points." \
  --max-tokens 512 \
  --temperature 0.0

Add --enable-thinking when explicit thinking mode is desired and supported by the installed MLX-VLM version.

Image

python -m mlx_vlm.generate \
  --model z4bfg/Qwen3.8-27B-MLX-8bit \
  --image /path/to/image.jpg \
  --prompt "Describe the visible objects, their relationships, and the scene." \
  --max-tokens 512 \
  --temperature 0.0

Python

from mlx_vlm import generate, load
from mlx_vlm.prompt_utils import apply_chat_template

model_id = "z4bfg/Qwen3.8-27B-MLX-8bit"
model, processor = load(model_id)

images = ["/path/to/image.jpg"]
prompt = apply_chat_template(
    processor,
    model.config,
    "Describe the visible objects and the scene.",
    num_images=len(images),
)

result = generate(
    model=model,
    processor=processor,
    prompt=prompt,
    image=images,
    max_tokens=512,
    temperature=0.0,
)
print(result.text)

For server usage, multi-image input, video handling, thinking controls, and streaming, follow the current mlx-vlm documentation.

Local validation

These are local smoke and regression checks, not official Qwen benchmarks and not a substitute for a broad accuracy evaluation.

The converted checkpoint and the exact upstream BF16 revision were run with deterministic decoding on the same Apple Silicon host and test harness.

Five of six generated outputs were byte-identical between the two builds. The remaining failed probability case produced the same incorrect answer in both builds, so that specific failure was not introduced by quantization. This small result does not establish general accuracy parity.

Additional deployment checks completed successfully:

• direct MLX-VLM text and single-image inference;
• strict JSON output and syntactically valid Python generation;
• Russian-language instruction following;
• LM Studio text, image, and required tool-call requests;
• repeated unload/load cycles with a stable model identifier;
• four parallel prediction slots and an eight-request queue on a 512 GB M3 Ultra host.

Performance depends on prompt length, output length, runtime version, thermal state, context allocation, and concurrent workloads. The figures above should not be treated as cross-platform benchmarks.

Limitations and responsible use

• Quantization can change logits and outputs. Validate this checkpoint on your own task before relying on it.
• The model can hallucinate, produce incorrect code or factual claims, reflect training-data biases, and generate unsafe or inappropriate content.
• Tool calls are model-generated data. Validate tool names and arguments, apply least-privilege permissions, and require confirmation for consequential actions.
• The advertised context length is an architectural capability, not a memory guarantee. KV-cache and media-token memory can dominate at long contexts.
• Text and image paths were validated for this release. Video artifacts are retained, but video inference was not independently validated as part of this conversion release.
• This repository does not reproduce the upstream benchmark suite. Use the upstream card for official benchmark results and evaluation methodology.

Do not use model output as the sole basis for medical, legal, financial, safety-critical, or other high-impact decisions.

Reproducibility and integrity

The conversion environment, command, source revision, and modified files are documented in MODIFICATIONS.md. File hashes are listed in SHA256SUMS.

License and attribution

The upstream checkpoint is distributed under the Apache License 2.0. This repository retains that license in LICENSE, preserves upstream attribution, identifies the modified artifacts, and uses the same apache-2.0 license metadata.

Copyright in the original model remains with its original rightsholders. The maintainer of this repository claims only responsibility for the conversion, packaging, and release documentation, not ownership of the upstream model or its trademarks.

See NOTICE and MODIFICATIONS.md for additional attribution and modification notices.

Citation

Please cite the upstream Qwen work when using this conversion:

@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}
}

Acknowledgements

• Qwen Team and Alibaba Cloud for the original model.
• Apple ML Research for MLX.
• The mlx-vlm maintainers and contributors for the conversion and inference tooling.
• LM Studio for one of the local compatibility and serving environments used during validation.