README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen3.6-35B-A3B
library_name: mlx
tags:
- mlx
- qwen3.6
- qwen3_5_moe
- apple-silicon
- tool-calling
- go
---

# Qwen3.6-35B-go-v2 4-bit MLX

A Go-focused Qwen3.6-35B-A3B model for Apple Silicon, packaged in MLX.

Use it as a coding assistant for Go projects: generating focused patches, explaining diffs, tightening tests, reading tool outputs, and making small repo-aware edits. It was tested with [Swival](https://swival.dev) on file-editing and command-running workflows.

This is the plain 4-bit compatibility variant. It does not include native MTP tensors, so it is the best starting point if your MLX loader does not support MTP sidecars.

## Which Variant Should I Use?

- **Use this repo** if you want the smallest plain MLX package or need a loader-friendly non-MTP model.
- Use [`jedisct1/Qwen3.6-35B-go-v2-8bit.mlx`](https://huggingface.co/jedisct1/Qwen3.6-35B-go-v2-8bit.mlx) if you want more precision without native MTP.
- Use [`jedisct1/Qwen3.6-35B-go-v2-bf16.mlx`](https://huggingface.co/jedisct1/Qwen3.6-35B-go-v2-bf16.mlx) if you want full precision without native MTP.
- Use [`jedisct1/Qwen3.6-35B-go-v2-MTP-4bit.mlx`](https://huggingface.co/jedisct1/Qwen3.6-35B-go-v2-MTP-4bit.mlx) if your runtime supports native MTP and you want the faster MTP path.

## Usage

Requires [mlx-lm](https://github.com/ml-explore/mlx-examples/tree/main/llms/mlx_lm):

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

```python
from mlx_lm import load, generate

model, tokenizer = load("jedisct1/Qwen3.6-35B-go-v2-4bit.mlx")

messages = [
    {"role": "system", "content": "You are an expert Go developer."},
    {"role": "user", "content": "Generate a focused patch that replaces the manual retry loop in fetchUser() with the shared retry helper."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=500)
print(response)
```

## What It Is Good At

- Writing idiomatic Go patches from a concise change request.
- Explaining Go diffs in commit-message style.
- Following tool-calling workflows where it needs to inspect files before editing.
- Keeping changes focused instead of turning small fixes into broad rewrites.
- Working with tests, command output, and repository context.

## Limitations

- Outputs should be reviewed before use, especially patches that touch production systems.
- The model works best on focused Go changes, tests, and explanations. Very large refactors may need to be split into smaller steps.
- Tool calling depends on the runtime and client preserving the chat template and tool schema format.