Vela 2.0 0.8B, MLX 8-bit

An 8-bit MLX conversion of vllm-sr/Vela-2.0-0.8B at revision 326b01d81f61fff7fd4ae570e54b87a49119e490.

• model.safetensors, config.json: the Qwen3.5 backbone, renamed to the mlx_lm qwen3_5 layout and quantized with mlx_lm 0.32.0 (affine, 8 bits, group size 64). mlx_lm moves the conv1d axis and adds 1 to the zero-centred RMSNorm weights, as it does for any Qwen3.5 checkpoint.
• head.safetensors: the original decision head (head.*) and set_bias, in FP32.
• vela/: the original config.json, calibration.json, tokenizer, LICENSE, NOTICE and ATTRIBUTIONS.md, unchanged.

Only the decision head is here. Span questions (PII, hallucination and open-label spans) need the span heads, which this repository does not include.

It is a local classifier of Mail Assistant, which runs it on mlx-swift 0.32.3.

Parity

Measured on 95 private emails with one Choice question of 19 options (prompts of 1.3k to 3.5k tokens, no truncation):

• Same answer as the official PyTorch FP32 runtime on 95 of 95 emails, with a maximum confidence difference of 0.025.
• Same answers as the Python MLX runner with the same weights on 95 of 95, with a maximum probability difference of 2e-6.

License

Apache-2.0, like the original model. The original LICENSE, NOTICE and ATTRIBUTIONS.md are in vela/. The backbone weights are renamed and quantized; the head is moved to its own file. Nothing else is changed.