Key facts
- Flint uses 2.5 billion parameters, making it small enough for a laptop.
- Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip.
- Recursive's Eureka Machine contributed to Flint's development from pre-training to inference.
AI startup Recursive launched its first model, Flint, on Wednesday. The company, which focuses on AI that can improve itself, stated that Flint was largely built with the help of its own AI system, the Eureka Machine. This development is an early test of recursive self-improvement, an idea that raises concerns about AI control but is currently limited by human oversight.
Recursive, which secured $650 million in funding at a $4.65 billion valuation in May, aims for its AI to build better AI over time. Flint is designed to run on a user's laptop, offering offline use and privacy, and is available for free under an open-source license.
The model uses a mixture-of-experts approach with 2.5 billion active parameters out of a total of 40 billion, making it significantly smaller and faster than some comparable models. In tests, Flint achieved speeds of approximately 131 tokens per second on a MacBook Pro M4 Pro, outperforming other models in its size class on software bug fixing benchmarks.