moonshotai/Kimi-K2.7-Code optimized for running on a Mac Studio M3 Ultra.
- A mixed-precision quant that balances speed, memory, and accuracy.
- 3-bit MoE baseline with important always-on layers at higher precision.
- Fits into ~460 GB memory, leaving enough room for a smaller utility model.
Usage
# Start server at http://localhost:8080/v1/chat/completions
uvx --from mlx-lm mlx_lm.server \
--host 127.0.0.1 \
--port 8080 \
--model spicyneuron/Kimi-K2.7-Code-MLX-3.6bit
Benchmarks
| metric | this model |
|---|---|
| bpw | 3.578 |
| base memory | 427.579 |
| peak memory (1024/512) | 460.444 |
| prompt tok/s (1024) | 218.851 ± 0.208 |
| gen tok/s (512) | 21.035 ± 0.049 |
| perplexity | 4.462 ± 0.037 |
| arc_challenge | 0.692 ± 0.021 |
| hellaswag | 0.780 ± 0.019 |
Methodology
Quantized with a mlx-lm fork. MLX quantization options differ than llama.cpp, but the principles are the same:
- Sensitive layers like MoE routing, attention, and output embeddings get higher precision
- More tolerant layers like MoE experts get lower precision