Kapitsa Flint
Flint 40B-A2.5B (4-bit, 22.6 GB) running locally on Apple Silicon through Kapitsa. Recursive's coding-agent MoE with reasoning and tool calls.
kapitsa serve --model flint
Speed
| Mac | Prefill (tok/s) | Decode (tok/s) |
|---|---|---|
| M5 Max 128 GB | 5,770 | 265 |
| M5 Pro 48 GB | 2,691 | 149 |
| M4 Pro 48 GB | 968 | 128 |
| M3 Pro 36 GB | 801 | 74 |
| M1 Pro 16 GB (2-bit, 11.9 GB) | 414 | 61 |
Architecture
Mixture of Experts, 40B total parameters, 2.5B active per token, 4-bit quantization.
Weights
Model tree for nace-ai/kapitsa-flint
Base model
RecursiveSI/flint-int4