Base model: Qwen3.5-35B-A3B with REAP expert pruning (20% pruned, 256 -> 205 experts)
Quantization: 4-bit affine, group size 64
Size on disk: ~16 GB
Peak memory: ~16 GB
Generation speed: ~64 tokens/sec (Apple M-series)

Usage

bash

pip install mlx-lm
mlx_lm.generate --model Alanfwilliams/Qwen-3.5-28B-A3B-REAP-4bit-mlx --prompt "Hello"

Or serve as an OpenAI-compatible API:

mlx_lm.server --model Alanfwilliams/Qwen-3.5-28B-A3B-REAP-4bit-mlx --port 8080

Conversion

Converted using a custom script that handles REAP's pruned expert layout. The standard mlx_lm.convert doesn't support REAP models because the per-expert weight keys need to be fused into MLX's SwitchGLU format with the correct (pruned) expert count.

Credits

Original model: 0xSero/Qwen-3.5-28B-A3B-REAP
REAP pruning method: arXiv:2510.13999
Base architecture: Qwen3.5-35B-A3B by Alibaba