---
title: Molmo2-8B-mxfp4
canonical_url: "https://www.modelscope.cn/models/mlx-community/Molmo2-8B-mxfp4"
md_url: "https://www.modelscope.cn/models/mlx-community/Molmo2-8B-mxfp4.md"
repository: mlx-community/Molmo2-8B-mxfp4
last_updated: 2026-10-20
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - molmo2
architectures:
  - Molmo2ForConditionalGeneration
base_model:
  - allenai/Molmo2-8B
base_model_relation: quantized
parameters: 420.6M
tensor_type:
  - F16
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 66
stars: 0
tags:
  - mlx
  - multimodal
  - olmo
  - molmo
  - molmo2
  - 4-bit
  - mxfp4
---

# Molmo2-8B-mxfp4

> Molmo2-8B-mxfp4 - mlx-community 在 ModelScope 开源的模型。This model was converted to MLX format from allenai/Molmo2-8B using mxlfp converter version 0.3.x.

mlx-community/Molmo2-8B-mxfp4 是 ModelScope 魔搭社区上的 420.6M 参数image-text-to-text模型，采用 apache-2.0 许可，基于 allenai/Molmo2-8B 构建。

- **Repository**: mlx-community/Molmo2-8B-mxfp4
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 420.6M
- **Base model**: allenai/Molmo2-8B
- **Tags**: mlx, multimodal, olmo, molmo, molmo2, 4-bit, mxfp4
- **Downloads**: 66
- **Stars**: 0
- **Last updated**: 2026-10-20

Source: https://www.modelscope.cn/models/mlx-community/Molmo2-8B-mxfp4

---

# Molmo2-8B-mxfp4

This model was converted to MLX format from
[`allenai/Molmo2-8B`](https://huggingface.co/allenai/Molmo2-8B) using
[mxlfp converter](https://github.com/Blaizzy/mlx-vlm) version **0.3.x**.

Refer to the [original model card](https://huggingface.co/allenai/Molmo2-8B)
for more details on the model.

## Quantization

This is a **4-bit MXFP4** conversion:

- Quantization mode: `mxfp4`
- Bits: `4`
- Group size: `32`
- Converted with dtype `float16`

MXFP4 is a 4-bit floating-point format that uses a shared power-of-two scale
per group of 32 weights. It differs from affine (integer) 4-bit quantization
and from NVIDIA's NVFP4 format.

The quantization is applied per-module; the vision tower and multimodal
projector are kept at higher precision, so the effective size is ~5.7 bits per
weight (5.8 GB on disk) rather than exactly 4 bits per parameter.

## Use with MLX

```sh
pip install mlx-vlm
```

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "mlx-community/Molmo2-8B-mxfp4"

model, processor = load(model_path)
config = load_config(model_path)

image = ["path/to/image.jpg"]
prompt = "Describe this image."

formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=len(image)
)

output = generate(model, processor, formatted_prompt, image, verbose=False)
print(output.text)
```

### CLI

```sh
mlx_vlm generate \
  --model mlx-community/Molmo2-8B-mxfp4 \
  --image path/to/image.jpg \
  --prompt "Describe this image." \
  --max-tokens 150
```