mlx-community/MAI-UI-8B-4bit

MAI-UI-8B 4bit

This is a 4-bit quantized MLX conversion of Tongyi-MAI/MAI-UI-8B, optimized for Apple Silicon.

MAI-UI is a family of real-world centric foundation GUI agents built for grounding, GUI navigation, user interaction, and broader device-cloud agent workflows. The family spans multiple scales and is framed upstream around realistic deployment, including user interaction, MCP-style tool use, online RL, and device-cloud collaboration.

This artifact was derived from the validated local MLX bf16 reference conversion and then quantized with mlx-vlm. It was validated locally with both mlx_vlm prompt-packet checks and vllm-mlx OpenAI-compatible serve checks.

Conversion Details

Field	Value
Upstream model	Tongyi-MAI/MAI-UI-8B
Artifact type	4bit quantized MLX conversion
Source artifact	local validated bf16 MLX artifact
Conversion tool	mlx_vlm.convert via mlx-vlm 0.3.12
Python	3.11.14
MLX	0.31.0
Transformers	5.2.0
Validation backend	vllm-mlx (phase/p1 @ 8a5d41b)
Quantization	4bit
Group size	64
Quantization mode	affine
Converter dtype note	float32
Reported effective bits per weight	6.776
Artifact size	6.93G
Template repair	tokenizer_config.json["chat_template"] was re-injected from chat_template.jinja after quantization

Additional notes:

• Root-level packaging is intentional for vllm-mlx multimodal detection compatibility.
• processor_config.json and video_preprocessor_config.json are present at repo root.
• This artifact intentionally augments tokenizer-visible template metadata for downstream compatibility checks.

Validation

This artifact passed local validation in this workspace:

• mlx_vlm prompt-packet validation: PASS
• vllm-mlx OpenAI-compatible serve validation: PASS

Local validation notes:

• output shape stayed aligned with the local bf16 and 6bit reference artifacts
• grounding drift increased relative to 6bit, but still returned the correct label and a plausible lower-screen input-region box
• the known baseline schema limitation remained unchanged from bf16: the structured-action output still omitted the requested reason field

Performance

• Artifact size on disk: 6.93G
• Local fixed-packet mlx_vlm validation used about 32.74 GB peak memory
• Observed local fixed-packet throughput was about 164-172 prompt tok/s and 44.5-48.7 generation tok/s across the four validation prompts
• Local vllm-mlx non-stream request time was about 27.46s, materially slower than the bf16 reference run and close to 6bit

These are local validation measurements, not a full benchmark suite.

Usage

Install

pip install -U mlx-vlm

CLI

python -m mlx_vlm.generate \
  --model mlx-community/MAI-UI-8B-4bit \
  --image path/to/image.png \
  --prompt "Describe the visible controls on this screen." \
  --max-tokens 256 \
  --temperature 0.0

Python

from mlx_vlm import load, generate

model, processor = load("mlx-community/MAI-UI-8B-4bit")
result = generate(
    model,
    processor,
    prompt="Describe the visible controls on this screen.",
    image="path/to/image.png",
    max_tokens=256,
    temp=0.0,
)
print(result.text)

vllm-mlx Serve

python -m vllm_mlx.cli serve mlx-community/MAI-UI-8B-4bit --mllm --localhost --port 8000

Links

• Upstream model: Tongyi-MAI/MAI-UI-8B
• Paper: MAI-UI Technical Report: Real-World Centric Foundation GUI Agents
• Project page: tongyi-mai.github.io/MAI-UI
• GitHub: Tongyi-MAI/MAI-UI
• MLX framework: ml-explore/mlx
• mlx-vlm: Blaizzy/mlx-vlm

Other Quantizations

Planned sibling repos in this wave:

• mlx-community/MAI-UI-8B-bf16
• mlx-community/MAI-UI-8B-6bit
• mlx-community/MAI-UI-8B-4bit - this model

Notes and Limitations

• This card reports local MLX conversion and validation results only.
• Upstream benchmark claims belong to the original MAI-UI model family and were not re-run here unless explicitly stated.
• Quantization changes numerical behavior relative to the local bf16 reference artifact.
• In local validation, the main trade relative to bf16 was increased grounding drift plus slower prefill, not response collapse.

Citation

If you use this MLX conversion, please also cite the original MAI-UI work:

@misc{zhou2025maiuitechnicalreportrealworld,
  title={MAI-UI Technical Report: Real-World Centric Foundation GUI Agents},
  author={Hanzhang Zhou and Xu Zhang and Panrong Tong and Jianan Zhang and Liangyu Chen and Quyu Kong and Chenglin Cai and Chen Liu and Yue Wang and Jingren Zhou and Steven Hoi},
  year={2025},
  eprint={2512.22047},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.22047},
}

License

This repo follows the upstream model license: Apache 2.0. See the upstream model card for the authoritative license details: Tongyi-MAI/MAI-UI-8B.