CoolFace
Modelpublic

ronnie3786/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-MLX-4bit

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
1likes63downloads
Model Card

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 (MLX 4-bit)

A 4-bit MLX quantization of Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2, optimized for Apple Silicon.

About the Base Model

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 is the second iteration of the reasoning-focused Qwen3.5-27B fine-tune. It's built on top of Qwen3.5-27B and fine-tuned with Unsloth using distilled reasoning data from Claude 4.6 Opus. This model has been #1 trending on HuggingFace for 3 weeks.

Key improvements in v2:

  • —More efficient chain-of-thought generation
  • —Improved reasoning across science, math, and instruction-following
  • —Trained on data from Jackrong/Qwen3.5-reasoning-700x

Why MLX?

MLX is Apple's native machine learning framework for Apple Silicon. On an M1 Ultra Mac Studio (128GB), this model runs at 28.9 tokens/sec compared to 19.9 tok/s for the equivalent GGUF Q4KM. That's a 45% speed improvement for free, just by using the right format for your hardware.

Benchmark Results

Controlled test: LM Studio, one model loaded at a time, 10 runs each with different prompts, 512 max tokens.

FormatAvg tok/sStd DevMinMax
MLX 4-bit28.90.1128.829.1
GGUF Q4KM19.90.0319.919.9

Hardware: M1 Ultra, 128GB unified memory.

How to Use

LM Studio

Download and place in your LM Studio models directory. LM Studio auto-detects MLX models.

mlx-lm CLI

bash
pip install mlx-lm
mlx_lm.generate --model ronnie3786/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-MLX-4bit --prompt "Your prompt here"

mlx-lm Server (OpenAI-compatible API)

bash
mlx_lm.server --model ronnie3786/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-MLX-4bit --port 8080

Conversion Details

  • —Quantization: 4-bit (4.501 bits per weight)
  • —Tool: mlx-lm 0.31.1 via mlx_lm.convert
  • —Size: ~14GB
  • —RAM required: ~16GB (fits comfortably on any Apple Silicon Mac with 16GB+ unified memory)
bash
mlx_lm.convert \
  --hf-path Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 \
  -q --q-bits 4 \
  --mlx-path ./output

Note: Requires mlx-lm >= 0.31.1. Earlier versions don't support the qwen3_5 architecture.

Credits