junhoyeo/Motif-2.6B-MLX-q4

Motif-2.6B — MLX 4-bit

4-bit MLX conversion of Motif-Technologies/Motif-2.6B for Apple Silicon, produced by mlx-motif — the MLX port of Motif's Differential Attention + PolyNorm architecture. This 2.6B variant uses the ungrouped ("vanilla") differential-attention form.

This checkpoint requires mlx-motif (it registers the model class into mlx-lm's loader); it will not load with stock mlx_lm.load.

Usage

git clone https://github.com/junhoyeo/mlx-motif && cd mlx-motif
uv pip install -e .

mlx-motif generate --model <this-repo> --prompt "Hello, world."
mlx-motif serve --model <this-repo> --port 8080   # OpenAI-compatible

Conversion provenance

• Converter: mlx-motif convert --hf-path Motif-Technologies/Motif-2.6B --out … --quantize --bits 4 (group_size 64, uniform preset)
• mlx-motif: github.com/junhoyeo/mlx-motif @ e6c401a (converted with this repo's convert.py; validated at this commit)
• mlx version: 0.31.2

Validation (measured on Apple M1 Max, 64 GB)

• End-to-end greedy generation verified on real weights (Python CLI, OpenAI server, and the native Swift runtime).
• Known numerical note: on this checkpoint, mlx-motif's custom-kernel path and its pure-MLX reference path produce output that diverges within a few greedy tokens — the two paths accumulate float reductions in different orders and the low-order-bit difference can flip the argmax on this architecture. Both outputs are valid samples of the model; this is documented (and expected) behavior, unlike the 12.7B grouped checkpoint where the two paths are byte-identical.
• Parity against the HF PyTorch reference is verified at bf16, not at q4.
• A q4 perplexity number for this checkpoint has not yet been recorded (the 12.7B's is 12.365); it will be added on the next benchmark pass.

License & attribution

The model weights are derivative of Motif Technologies' release and remain under Apache 2.0, © Motif Technologies. The conversion tooling is MIT (mlx-motif). If you use this model, please attribute the original model card.