junhoyeo/Motif-2-12.7B-Reasoning-MLX-q4

Motif-2-12.7B-Reasoning — MLX 4-bit

4-bit MLX conversion of Motif-Technologies/Motif-2-12.7B-Reasoning for Apple Silicon, produced by mlx-motif — the MLX port of Motif's Grouped Differential Attention + PolyNorm architecture, with custom Metal kernels for the differential-attention decode path.

Neither architecture primitive ships in mlx-lm, so 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

from mlx_lm import generate
from mlx_motif import load

model, tokenizer = load("<this-repo>")
print(generate(model, tokenizer, prompt="…", max_tokens=128))

Conversion provenance

• Converter: mlx-motif convert --hf-path Motif-Technologies/Motif-2-12.7B-Reasoning --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)

• Perplexity: 12.365 (nll/token 2.5149, 592 tokens, scripts/perplexity.py --chunk 512).
• Kernel-path parity: greedy output on this checkpoint is byte-identical between mlx-motif's custom Metal kernels and its pure-MLX reference path (32/48/96-token checks on real weights). Parity against the HF PyTorch reference is verified at bf16, not at q4 — 4-bit quality is characterized by the perplexity above instead.
• Decode throughput (July 2026, bench_decode_e2e.py, median of 5 runs, max_tokens=64, default configuration):

Prompt length	tok/s
5	40.9
164	40.0
800	38.2
3204	30.7

Notes & limitations

• Motif's chat template pre-opens a <think> reasoning block; mlx-motif serve --think-mode visible|hidden|captured controls how the trace is surfaced.
• Kernel constants are tuned on M1 Max; other M-series chips work but have not been performance-validated.
• Quantization: uniform 4-bit, group size 64. Mixed-precision presets are available in the converter if you want a different quality/size point.

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.