Kapitsa 35B
Qwen3.6 35B-A3B MoE (4-bit, 22.1 GB) running locally on Apple Silicon through Kapitsa. Fast replies and fast reading.
kapitsa serve --model 35b
Speed
| Mac | Prefill (tok/s) | Decode (tok/s) |
|---|---|---|
| M5 Max 128 GB | 6,539 | 351 |
| M5 Pro 48 GB | 2,802 | 201 |
| M4 Pro 48 GB | 1,168 | 161 |
| M3 Pro 36 GB | 855 | 106 |
| M1 Pro 16 GB (2-bit, 10.8 GB) | 462 | 52 |
Architecture
Mixture of Experts, 35B total parameters, 3B active per token, 4-bit quantization.