nace-ai/kapitsa-9b

Kapitsa 9B

Qwen3.5 9B (4-bit, 9.0 GB) running locally on Apple Silicon through Kapitsa.

kapitsa serve --model 9b

Speed

Mac	Prefill (tok/s)	Decode (tok/s)
M3 Pro 36 GB	381	77
M1 Pro 16 GB	217	57

Architecture

Dense transformer, 9B parameters, 4-bit quantization.

Weights

lmstudio-community/Qwen3.5-9B-MLX-4bit