Recursive SI released Flint, a 40-billion-parameter mixture-of-experts coding model, on October 21, with weights on Hugging Face at RecursiveSI/flint under Apache 2.0 — no revenue gate, no OEM clause, no service-operator trigger, per Forkast News.

OrcaRouter reports the release is actually four repositories: a BF16 checkpoint of 79.59GB plus int4 safetensors, int4 GGUF and NVFP4 GGUF builds, the last at just 22.86GB. The company says the model was built in under four months (company-reported).

The number that defines Flint is not its total size but its active-parameter ratio: 39.8 billion total parameters with roughly 2.5 billion active per token, according to OrcaRouter.

With 256 routed experts per MoE layer and 6 selected per token, decoding reads almost none of the weights, which is why a laptop can run it fast.

OrcaRouter reports Recursive's Eureka Machine — an automated research loop that proposes, implements and tests architectures — searched expert configurations from a 35B baseline and projected this 40B/2.5B shape as the best speed-versus-coding tradeoff among its candidates.

Forkast News reports Recursive claims more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark — vendor-reported, measured as average tokens per second across a 128,000-token session with prefill excluded.

OrcaRouter notes Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, all vendor-reported and unreproduced; no independent index had a Flint entry as of October 22.

The evaluation protocol averages pass@1 over at least five attempts with web-retrieval solutions counted as failures, per Forkast News.

One specification does not match across the release surfaces: the announcement describes a 512,000-token context window with 256,000 recommended, while the Hugging Face model card documents 262,144 tokens, according to Forkast News.

The discrepancy is stated rather than resolved because neither surface has been updated to reconcile it.

OrcaRouter adds the model card is explicit that the 262,144 figure covers prompt and generated output together, and that the same vendor advertises "up to 512k" one paragraph from recommending 256k for coding.

The Apache 2.0 license places Flint in the unrestricted tier — the same class as DeepSeek's DeepInfer engine and NVIDIA's Nemotron Retriever models, per Forkast News — letting developers download, modify, redistribute and deploy commercially at any scale.

That makes Flint the weights-to-watts argument at its smallest unit: a coding model sized for a single developer machine with no growth gate and no cloud dependency.

OrcaRouter adds the quantised output layer and custom Metal megakernel show Recursive controls architecture, quantisation and kernels together, and its benchmarks target local peers like Qwen 3.8 and Gemma 4 rather than frontier closed models.