Alice-Qwen3.5-9B-Code-v0-MLX-8bit

8-bit MLX build for the Alice 9B code/security lane.

Files

• model-00001-of-00002.safetensors
• model-00002-of-00002.safetensors
• tokenizer and config files

Source

Converted from local HF fused BF16:

• /Users/v/Documents/New project/models/hf/Alice-Qwen3.5-9B-Code-v0-HF-Fused-bf16

Quantization:

• MLX affine 8-bit
• group size 64
• effective bits per weight: about 8.501

Alice Profile

The model folder includes the Alice code/security chat template and corrected EOS handling for <|im_end|>.

Target use:

• Solidity
• TypeScript and Python utility work
• Foundry / Hardhat workflows
• smart-contract security review
• concise Chinese/English engineering answers

Local Smoke

Prompt:

你是谁？用一句话说你擅长什么。

Output:

我是 Alice，擅长用 Solidity、TypeScript 和 Python 写可审计合约、补丁漏洞、写测试和工程落地清单。

Observed on local MLX:

• Prompt: about 220 tok/s
• Generation: about 37 tok/s
• Peak memory: about 9.83 GB