---
library_name: mlx
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/LICENSE
pipeline_tag: text-generation
tags:
- mlx
base_model: Qwen/Qwen3.5-397B-A17B
---

# m-i/Qwen3.5-397B-A17B-Text-2.423bit [text only]

This model [m-i/Qwen3.5-397B-A17B-Text-2.423bit](https://huggingface.co/m-i/Qwen3.5-397B-A17B-Text-2.423bit) was
converted to MLX format from [Qwen/Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)
using mlx-lm version **0.31.0**.

## Parameters

Try lower temp than the one recommended for full precision. --temp 0.5 or lower.

## quant predicate

```python
def qwen397b_predicate(path: str, module, ):
    # MLP projection layers are typically largest and most robust to aggressive quantization
    if any(proj in path for proj in ["down_proj"]):
        return {"group_size": 64, "bits": 2, "mode": "affine"}
    if any(proj in path for proj in [ "up_proj", "gate_proj"]):
        return {"group_size": 128, "bits": 2, "mode": "affine"}

    if "lm_head" in path:
        return {"group_size": 128, "bits": 6, "mode": "affine"}

    if "embed_tokens" in path:
        return {"group_size": 128, "bits": 8, "mode": "affine"}

    # All other weights: attention projections, norms, etc.
    return {"group_size": 32, "bits": 5, "mode": "affine"}
```

## Use with mlx

```bash
pip install mlx-lm
```

```python
from mlx_lm import load, generate

model, tokenizer = load("m-i/Qwen3.5-397B-A17B-Text-2.423bit")

prompt = "hello"

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

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