---
title: GLM-5.3-mixed-4-8
canonical_url: "https://www.modelscope.cn/models/mlx-community/GLM-5.3-mixed-4-8"
md_url: "https://www.modelscope.cn/models/mlx-community/GLM-5.3-mixed-4-8.md"
repository: mlx-community/GLM-5.3-mixed-4-8
last_updated: 2026-08-30
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - glm_moe_dsa
architectures:
  - GlmMoeDsaForCausalLM
base_model:
  - zai-org/GLM-5.3
base_model_relation: quantized
parameters: 118.6B
tensor_type:
  - BF16
  - U32
  - F32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
  - zh
downloads: 11
stars: 0
tags:
  - mlx
  - transformers
---

# GLM-5.3-mixed-4-8

> GLM-5.3-mixed-4-8 - mlx-community 在 ModelScope 开源的模型。mlx-community/GLM-5.3-mixed-4-8

mlx-community/GLM-5.3-mixed-4-8 是 ModelScope 魔搭社区上的 118.6B 参数text-generation模型，采用 other 许可，基于 zai-org/GLM-5.3 构建。

- **Repository**: mlx-community/GLM-5.3-mixed-4-8
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 118.6B
- **Base model**: zai-org/GLM-5.3
- **Tags**: mlx, transformers
- **Downloads**: 11
- **Stars**: 0
- **Last updated**: 2026-08-30

Source: https://www.modelscope.cn/models/mlx-community/GLM-5.3-mixed-4-8

---

# mlx-community/GLM-5.3-mixed-4-8

This model [mlx-community/GLM-5.3-mixed-4-8](https://huggingface.co/mlx-community/GLM-5.3-mixed-4-8) was converted to MLX format from [zai-org/GLM-5.3-BF16](https://huggingface.co/zai-org/GLM-5.3-BF16) using mlx-lm version **0.31.3** (with PR #1410).

Note that this quant is using the GLM-5.3-BF16 as base.

This is created for people using a single Apple Mac Studio M3 Ultra with 512 GB. The 4-bit version of GLM-5.3 fits comfortably. But we can do better. We aim to get better results from a slightly larger and smarter quantization. It should also not be so large that it leaves no memory for a useful context window.

You can find more similar MLX model quants for Apple Mac Studio with 512 GB at https://huggingface.co/bibproj

```bash
pip install mlx-lm

mlx_lm.generate --model mlx-community/GLM-5.3-mixed-4-8 --prompt "Hi"
```

---

## What is this mixed-4-8?

This is "similar" to the recipe that `antirez` uses for his well-known DeepSeek-V4 / DwarfStar inference code.

---

## Recipe

To make to the quant perform better under stress, only the expert tensors are quantized to 4-bit. All the other tensors are kept at 8-bit. You could say that this quant has an 8-bit "brain" and 4-bit experts.

In the `convert.py` file of mlx-lm on your system ( [you can see the original code here](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/convert.py) ), replace the code inside `def mixed_quant_predicate()` with something like

```python
        # Build a mixed quant like "mixed-4-8"
        q_bits = 8
        if "switch_mlp" in path:
           q_bits = 4
        print("path:", path, "index:", index, "q_bits:", q_bits)
        return {"group_size": group_size, "bits": q_bits, "mode": mode}
```

Then create your GLM-5.3-mixed-4-8 quant with

```bash
mlx_lm.convert --hf-path zai-org/GLM-5.3-BF16 --mlx-path GLM-5.3-mixed-4-8 -q --quant-predicate mixed_4_6
```

---

Enjoy!