---
title: Muse-Spark-Open-816B-A42B-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/Muse-Spark-Open-816B-A42B-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/Muse-Spark-Open-816B-A42B-4bit.md"
repository: mlx-community/Muse-Spark-Open-816B-A42B-4bit
last_updated: 2026-10-18
license: other
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - muse_spark
architectures:
  - MuseSparkForConditionalGeneration
base_model:
  - meta-models/Muse-Spark-Open-816B-A42B
base_model_relation: quantized
parameters: 169.6B
tensor_type:
  - F32
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 495
stars: 0
tags:
  - mlx
---

# Muse-Spark-Open-816B-A42B-4bit

> Muse-Spark-Open-816B-A42B-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/Muse-Spark-Open-816B-A42B-4bit

mlx-community/Muse-Spark-Open-816B-A42B-4bit 是 ModelScope 魔搭社区上的 169.6B 参数image-text-to-text模型，采用 other 许可，基于 meta-models/Muse-Spark-Open-816B-A42B 构建。

- **Repository**: mlx-community/Muse-Spark-Open-816B-A42B-4bit
- **License**: other
- **Tasks**: image-text-to-text
- **Parameters**: 169.6B
- **Base model**: meta-models/Muse-Spark-Open-816B-A42B
- **Tags**: mlx
- **Downloads**: 495
- **Stars**: 0
- **Last updated**: 2026-10-18

Source: https://www.modelscope.cn/models/mlx-community/Muse-Spark-Open-816B-A42B-4bit

---

# mlx-community/Muse-Spark-Open-816B-A42B-4bit

This model [mlx-community/Muse-Spark-Open-816B-A42B-4bit](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-4bit) was
converted to MLX format from [meta-models/Muse-Spark-Open-816B-A42B](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)
using mlx-lm version **0.32.2**.

Re-uploaded on 2026-10-16 at 07:21 UTC with mlx-lm 0.32.2, which supports Muse Spark.

mlx-lm's default quantization for this model makes the routed experts of layers 1-60 4-bit and keeps the other weights in BF16. The checkpoint includes the vision encoder and takes about 533 GB, more than the memory of a single 512 GB Mac.

## Use with mlx

```bash
pip install -U "mlx-lm>=0.32.2"
```

```python
from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Muse-Spark-Open-816B-A42B-4bit")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)
```

## You can also use

- [mlx-community/Muse-Spark-Open-816B-A42B-oQ3e](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-oQ3e): mixed-precision oQ quant optimized for [oMLX](https://github.com/jundot/omlx), with the vision encoder. It also runs in mlx-lm 0.32.2.
- [mlx-community/Muse-Spark-Open-816B-A42B-oQ4e](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-oQ4e): the 4-bit oQ quant.
- [mlx-community/Muse-Spark-Open-816B-A42B-3bit](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-3bit): the standard mlx-lm 3-bit quant.