The Flint model improved the performance of NanoChat across more than 700 machine learning trials by autonomously managing the research loop. Recursive reported that these results finished close to those of Claude Haiku 4.5, demonstrating that a model designed for laptop use can conduct complex training research typically associated with larger proprietary systems. The company is now testing how this self-improvement cycle can be expanded while maintaining safety protocols.
The 40B parameter mixture-of-experts model, which is’ activated by 2.5B parameters per token, is seeing wide deployment on local hardware. User benchmarks recorded 95 tokens per second on an Nvidia GB10, while other tests on an M4 Max showed Flint running 5x faster than Qwen 27B. The model, created in under four months via the Eureka Machine, is now available in GGUF formats that enable execution on devices with 24GB of RAM.