AI BriefingKapitsa Achieves 128 tok/s on MacBook Pro M4 Pro for Recursive's Flint 40B A2.5B Q4
·2026.10.22 21:17
Key point
Open-source engine Kapitsa matches Recursive's native performance on short contexts and leads by 1.30x at 128K context on Apple Silicon.
Details
- Kapitsa, an open-source inference engine for Apple Silicon, achieves 127.9 tok/s decode speed on a MacBook Pro M4 Pro (48GB) running Recursive's Flint 40B A2.5B Q4 model.
- Performance matches Recursive's native engine on short contexts (1% difference) and outperforms llama.cpp by 1.6x in decode speed.
- At 128K context, Kapitsa decode speed is 1.30x faster than Recursive's engine and 1.9x faster than llama.cpp.
- A KFlash tree drafter for speculative decoding is planned for release, targeting a 1.4x to 1.5x speedup in decode without accuracy loss.
- Benchmarks were conducted on macOS 27.0.1 with High Power mode; peak memory usage was 23.3 GB for Kapitsa compared to 27.9 GB for Recursive's engine.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.