AI startup Recursive has introduced Flint, an open-source Mixture of Experts (MoE) language model specifically tailored for local coding agents. This model impresses with approximately 40 billion total parameters, activating around 2.5 billion during runtime. It scored an outstanding 89.7% on the SWE-bench Pro v2 test, accommodates contexts up to 256K tokens, and exhibits remarkable inference speed on consumer-grade hardware, coupled with exceptional performance across various benchmarks.

Flint's development was significantly bolstered by Recursive's AI research system, Eureka Machine, which made pivotal contributions in three key areas: architecture search, quantization, and GPU kernel optimization. Remarkably, the model was trained from scratch within a span of less than four months.

Flint is distributed under the Apache 2.0 license, enabling individuals and enterprises to utilize, modify, and commercialize it without any cost. This accomplishment underscores the substantial role of AI in model development. However, human intervention remains essential for setting goals and guiding the direction, suggesting that complete autonomous recursive self-improvement is still an aspiration for the future.