Recursive Flint drops AI-built coding model

Flint is more interesting as a systems story than a parameter-count story: Recursive is pairing sparse modeling, hardware-specific optimization, and automated research to make serious coding agents viable on consumer hardware.

• –The MoE design keeps compute low while retaining a 40B-class parameter pool, helping Flint fit local machines through quantization.
• –Its long context and multi-turn tool use target repository-scale work—bug investigation, implementation, and testing—rather than simple code completion.
• –Apache 2.0 licensing and local execution eliminate API bills while keeping source code and prompts on-device.
• –Recursive says its Eureka Machine automated much of the model-development loop, including architecture, training, and inference improvements.
• –Benchmark claims are promising but publisher-reported, so independent evaluations will determine whether Flint is genuinely competitive with established coding models.