Kapitsa 27B
Qwen3.8 27B (4-bit, 20.2 GB) running locally on Apple Silicon through Kapitsa.
kapitsa serve --model 27b
Speed
| Mac | Prefill (tok/s) | Decode (tok/s) |
|---|---|---|
| M5 Max 128 GB | 1,126 | 144 |
| M5 Pro 48 GB | 606 | 77 |
| M4 Pro 48 GB | 396 | 63 |
| M3 Pro 36 GB | 110 | 40 |
Architecture
Dense transformer, 27B parameters, 4-bit quantization.
Weights
Model tree for nace-ai/kapitsa-27b
Base model
Qwen/Qwen3.8-27B Quantized
lmstudio-community/Qwen3.8-27B-MLX-4bit