qybaihe/mude-1.0-mlx-8bit

mude 1.0 (MLX 8-bit, Apple silicon)

This is the Mac edition of qybaihe/mude-1.0, a 2B-parameter decision model optimized for coding work and the decision model ("judge") of the coding agent mu. It contains:

• the same weights in MLX's layout, quantized to 8 bits (mlx-q8.safetensors, 1.76 GB);
• the decision head in bf16 (readout.safetensors).

The main model card covers what the model does, its evaluation, training data and limitations.

First published as qybaihe/mude-2b-mlx-8bit; that address redirects here, and the weights are unchanged.

中文：mude 1.0 是针对编程场景特别优化的决策模型，2B 参数；这里是它的 Mac 版（MLX 8 位）。原仓库名为 qybaihe/mude-2b-mlx-8bit，旧地址会自动跳转到这里。

Requirements

• An Apple-silicon Mac (M1 or newer)
• macOS 14 or newer
• Python 3.10–3.14

The compute dtype is chosen per chip: float16 on M1/M2, bfloat16 from M3 on. MLX's buffer cache is 1 GB on Macs with 16 GB or more and 0.25 GB below; D3_MLX_CACHE_GB overrides it.

Run

python3 -m venv .venv && .venv/bin/pip install -r requirements-mac.txt
export MUDE_KEY=$(python3 -c "import secrets; print(secrets.token_hex(16))")
DECISION_API_KEY=$MUDE_KEY .venv/bin/python d3_server_mac.py --model . --port 8790 --name mude

The server speaks the same System One API as the main edition's d3_server.py. Point mu at it with the mu.json shown in the main card.

From Python:

from d3_mac import D3Mac

model = D3Mac.from_pretrained(".")

Speed and parity

Measured on a Mac mini on a sample of 242 mu requests:

• 0.22 s per decision on average, 0.39 s at p90;
• 3.9 GB peak process memory;
• on 1048 dev questions, 3 are decided differently from the bf16 CUDA edition.

License

The weights and the Mac-specific files (d3_mac.py, d3_mlx.py, d3_server_mac.py) are released under MIT. Files taken unchanged from d3-nano and Qwen/Qwen3.5-2B stay under Apache-2.0. See NOTICE and LICENSES/Apache-2.0.txt.