---
title: UI-Venus-1.5-2B-6bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/UI-Venus-1.5-2B-6bit"
md_url: "https://www.modelscope.cn/models/mlx-community/UI-Venus-1.5-2B-6bit.md"
repository: mlx-community/UI-Venus-1.5-2B-6bit
last_updated: 2026-03-20
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen3_vl
architectures:
  - Qwen3VLForConditionalGeneration
base_model:
  - inclusionAI/UI-Venus-1.5-2B
base_model_relation: quantized
parameters: 851.5M
tensor_type:
  - U32
  - BF16
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 10
stars: 0
tags:
  - mlx
  - mlx-vlm
  - safetensors
  - apple-silicon
  - conversational
  - gui
  - vision-language-model
  - qwen3_vl
  - ui-venus
  - grounding
  - mobile-navigation
  - browser-navigation
  - computer-use
  - 6-bit
  - quantized
---

# UI-Venus-1.5-2B-6bit

> UI-Venus-1.5-2B-6bit - mlx-community 在 ModelScope 开源的模型。UI-Venus-1.5-2B 6bit

mlx-community/UI-Venus-1.5-2B-6bit 是 ModelScope 魔搭社区上的 851.5M 参数image-text-to-text模型，采用 apache-2.0 许可，基于 inclusionAI/UI-Venus-1.5-2B 构建。

- **Repository**: mlx-community/UI-Venus-1.5-2B-6bit
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 851.5M
- **Base model**: inclusionAI/UI-Venus-1.5-2B
- **Tags**: mlx, mlx-vlm, safetensors, apple-silicon, conversational, gui, vision-language-model, qwen3_vl, ui-venus, grounding, mobile-navigation, browser-navigation, computer-use, 6-bit, quantized
- **Downloads**: 10
- **Stars**: 0
- **Last updated**: 2026-03-20

Source: https://www.modelscope.cn/models/mlx-community/UI-Venus-1.5-2B-6bit

---

# UI-Venus-1.5-2B 6bit

This is a 6-bit quantized MLX conversion of [inclusionAI/UI-Venus-1.5-2B](https://huggingface.co/inclusionAI/UI-Venus-1.5-2B), optimized for Apple Silicon.

UI-Venus-1.5 is a unified end-to-end GUI agent family built for grounding, web navigation, and mobile navigation. The `1.5` family spans dense `2B` and `8B` variants plus a `30B-A3B` MoE variant, and is framed upstream around a shared GUI semantics stage, online RL for long-horizon navigation, and model merging across grounding, web, and mobile domains.

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 | `inclusionAI/UI-Venus-1.5-2B` |
| Artifact type | `6bit 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 | `6bit` |
| Group size | `64` |
| Quantization mode | `affine` |
| Converter dtype note | `float16` |
| Reported effective bits per weight | `8.086` |
| Artifact size | `2.31G` |
| Template repair | `tokenizer_config.json["chat_template"]` was re-injected after quantization |
| Inherited config workaround | local `bf16` base was created from an upstream mirror with `tie_word_embeddings = false` |

Additional notes:

- This MLX artifact preserves the dual-template contract across `chat_template.json`, `chat_template.jinja`, and `tokenizer_config.json["chat_template"]`.
- `chat_template.jinja` is present as an additive compatibility shim.
- This quantized artifact inherits the untied-embedding config from the validated local `bf16` base artifact.

## 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:

- the model stayed in the same weak-but-stable instruction-following envelope as the local `bf16` reference artifact
- the same action choice and the same weaker `Default Models` answer were preserved
- there was no new regression in the serve path; the main limitation remains model quality rather than packaging or runtime stability

## Performance

- Artifact size on disk: `2.31G`
- Local fixed-packet `mlx_vlm` validation used about `3.99 GB` peak memory
- Observed local fixed-packet throughput was about `536-565` prompt tok/s and `114.6-180.2` generation tok/s across the four validation prompts
- Local `vllm-mlx` serve validation completed in about `8.94s` non-stream and `9.68s` streamed

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

## Usage

### Install

```bash
pip install -U mlx-vlm
```

### CLI

```bash
python -m mlx_vlm.generate \
  --model mlx-community/UI-Venus-1.5-2B-6bit \
  --image path/to/image.png \
  --prompt "Describe the visible controls on this screen." \
  --max-tokens 256 \
  --temperature 0.0
```

### Python

```python
from mlx_vlm import load, generate

model, processor = load("mlx-community/UI-Venus-1.5-2B-6bit")
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

```bash
python -m vllm_mlx.cli serve mlx-community/UI-Venus-1.5-2B-6bit --mllm --localhost --port 8000
```

## Links

- Upstream model: [inclusionAI/UI-Venus-1.5-2B](https://huggingface.co/inclusionAI/UI-Venus-1.5-2B)
- Paper: [UI-Venus-1.5 Technical Report](https://arxiv.org/abs/2602.09082)
- Paper: [UI-Venus Technical Report: Building High-performance UI Agents with RFT](https://arxiv.org/abs/2508.10833)
- GitHub: [inclusionAI/UI-Venus](https://github.com/inclusionAI/UI-Venus)
- MLX framework: [ml-explore/mlx](https://github.com/ml-explore/mlx)
- mlx-vlm: [Blaizzy/mlx-vlm](https://github.com/Blaizzy/mlx-vlm)

## Other Quantizations

Planned sibling repos in this wave:

- [`mlx-community/UI-Venus-1.5-2B-bf16`](https://huggingface.co/mlx-community/UI-Venus-1.5-2B-bf16)
- [`mlx-community/UI-Venus-1.5-2B-6bit`](https://huggingface.co/mlx-community/UI-Venus-1.5-2B-6bit) - this model

## Notes and Limitations

- This card reports local MLX conversion and validation results only.
- Upstream benchmark claims belong to the original UI-Venus model family and were not re-run here unless explicitly stated.
- Quantization changes numerical behavior relative to the local `bf16` reference artifact.
- This artifact remains materially weaker than the UI-Venus `8B` family on precise UI instruction-following, even though the serve path stayed stable.

## Citation

If you use this MLX conversion, please cite the original UI-Venus papers:

- [UI-Venus-1.5 Technical Report](https://arxiv.org/abs/2602.09082)
- [UI-Venus Technical Report: Building High-performance UI Agents with RFT](https://arxiv.org/abs/2508.10833)

## License

This repo follows the upstream model license: Apache 2.0.
See the upstream model card for the authoritative license details:
[inclusionAI/UI-Venus-1.5-2B](https://huggingface.co/inclusionAI/UI-Venus-1.5-2B).