Flint open-weights coding model: Cline and Recursive dispute who released it

Recursive has released Flint, a 40B mixture-of-experts coding model under the Apache 2.0 license, according to Testing Catalog.

Recursive says Flint runs at more than 220 tokens per second on a 64GB RTX Spark, and that its Eureka Machine wrote GPU code that sped Flint up 55% on that hardware without changing the weights. Recursive says the weights are on Hugging Face.

Earlier, Cline said on X that Flint scores 54.9% on Terminal-Bench 2.1 with about 2B active parameters, and that it needs about 24GB of memory and runs at 125+ tokens per second on a MacBook Pro. Those figures are Cline's own claims, and no other source has confirmed the score. Thomas Wolf and Richard Socher, posting on X, attribute Flint to Recursive, describing it as Recursive's first model, built in under four months, with 4-bit lm_head quantization landing 0.7 points from bf16. The reports do not explain why Cline and Recursive are named as the source.