README.md

metadata

language:
  - en
library_name: mlx
tags:
  - qwen-coder
  - MOE
  - pruning
  - compression
  - mlx
license: apache-2.0
name: cerebras/Qwen3-Coder-REAP-25B-A3B
description: >
  This model was obtained by uniformly pruning 20% of experts in
  Qwen3-Coder-30B-A3B-Instruct using the REAP method.
readme: |
  https://huggingface.co/cerebras/Qwen3-Coder-REAP-25B-A3B/main/README.md
license_link: https://huggingface.co/cerebras/Qwen3-Coder-REAP-25B-A3B/blob/main/LICENSE
pipeline_tag: text-generation
base_model: cerebras/Qwen3-Coder-REAP-25B-A3B

Qwen3-Coder-REAP-25B-A3B-qx64-hi-mlx

The regular Deckard(qx) formula uses embeddings at the same bit as the data stores, in this case 4 bit.

The head and select attention paths are enhanced to 6 bit, and the model is quantized with group size 32(hi).

There is an updated model: Qwen3-Coder-REAP-25B-A3B-qx65x-hi-mlx that uses embeddings at 6 bit and a base of 5 bit, and should perform slightly better on long context.

Metrics coming soon.

-G

This model Qwen3-Coder-REAP-25B-A3B-qx64-hi-mlx was converted to MLX format from cerebras/Qwen3-Coder-REAP-25B-A3B using mlx-lm version 0.28.3.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Qwen3-Coder-REAP-25B-A3B-qx64-hi-mlx")

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)