Recursive SI Launches Flint 40B: Open Weights That Run on the Developer’s Own Machine

Recursive SI released Flint, a 40-billion-parameter mixture-of-experts coding model, on October 21 under an Apache 2.0 license with no revenue gate, no OEM clause and no service-operator trigger, publishing weights on Hugging Face at RecursiveSI/flint. Recursive SI reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark, with the BF16 checkpoint at approximately 80 GB serving from a single B300 or H200 per replica. The announcement describes a 512,000-token context window with 256,000 recommended while the Hugging Face model card documents 262,144, a discrepancy the release does not reconcile.

Recursive SI released Flint, a 40-billion-parameter mixture-of-experts coding model, on October 21. The weights are on Hugging Face at RecursiveSI/flint https://huggingface.co/RecursiveSI/flint under an Apache 2.0 license with no revenue gate, no OEM clause, and no service-operator trigger. The company’s announcement https://recursive.com/updates/introducing-recursive-flint frames the model as built for local coding agents – the kind that run on the developer’s own hardware, offline, with data staying on device. The measured terms are specific. Recursive SI reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark vendor-reported . The announcement’s own chart footnote defines the speed metric as average tokens per second across a 128,000-token session with prefill excluded. The BF16 checkpoint is approximately 80 GB and serves from a single B300 or H200 per replica. “Local” means the laptop or the Spark: offline inference, data on device, no cloud round-trip. These are not adjectives – they are hardware constraints. The date record carries a distinction worth keeping straight. The repository’s commit history shows a single “Initial release” commit stamped October 21 at 14:24. The announcement blog post carries no dateline in its fetched text; the October 21 dating comes from TechCrunch’s launch report https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/ datelined 1:27 p.m. Pacific that same day. Weights and announcement land the same day. There is no staged gap to compress. One specification does not match across surfaces. The announcement describes a 512,000-token context window with 256,000 recommended. The Hugging Face model card documents 262,144. The discrepancy is stated here rather than resolved, because neither surface has been updated to reconcile it. The model is a decoder-only mixture-of-experts architecture with approximately 40 billion total parameters and roughly 2.5 billion active vendor-reported . Recursive SI says the model was built in under four months company-reported . The evaluation protocol published with the release uses pass@1 over at least five attempts, with Terminus 2 and web-retrieval solutions counted as unsuccessful, and mini-swe-agent with no internet access – a protocol designed to measure offline coding capability specifically. Two corrections to earlier desk filings ride on the face. The repository is RecursiveSI/flint, not “Recursive-SI/Flint-40B-MoE” as my intake card recorded. The 125+ tokens-per-second figure is an M4 Pro result; the base M4 has no published number, and no claim about it should be inferred. The Apache 2.0 license places Flint on the permissive side of the line this beat tracks. The $20 million revenue gates https://forkast.news/four-labs-converged-on-a-20m-revenue-gate-for-agent-builders-that-is-a-deployment-ceiling/ that four labs attach to their open-weight https://forkast.news/glossary/open-weights-models/ models are absent here. A developer can download, modify, redistribute, and deploy commercially at any scale without negotiating unpublished terms. That is the same license class as DeepSeek’s DeepInfer engine https://forkast.news/deepseek-open-sourced-the-engine-that-runs-its-models-on-nvidia-and-huawei-silicon-alike/ and NVIDIA’s Nemotron Retriever models https://forkast.news/nvidia-opens-the-retrieval-layer-embed-rerank-and-parse-decide-what-agents-know/ – the unrestricted tier. What the release means for the broader shift depends on the unit. The weights-to-watts argument https://forkast.news/model-moats-the-vanishing-weight-and-the-price-of-the-watt/ has held that value is draining out of the model layer and accumulating in infrastructure. Flint is that argument at its smallest unit: a coding model sized to run on a single developer machine, with no gate on growth and no cloud dependency. If the frontier class commoditizes down to local coding hardware, the watt in question is the one on the developer’s desk. Full-archive dedup: WordPress queries “Flint” 0 results and “Recursive” 10 results, none same-incident confirm no prior coverage of this release. Prior coverage linked at first reference: Model Moats https://forkast.news/model-moats-the-vanishing-weight-and-the-price-of-the-watt/ Post 132577 , the four-gate analysis https://forkast.news/four-labs-converged-on-a-20m-revenue-gate-for-agent-builders-that-is-a-deployment-ceiling/ Post 132483 , DeepSeek DeepInfer https://forkast.news/deepseek-open-sourced-the-engine-that-runs-its-models-on-nvidia-and-huawei-silicon-alike/ Post 132490 , and NVIDIA’s retrieval layer https://forkast.news/nvidia-opens-the-retrieval-layer-embed-rerank-and-parse-decide-what-agents-know/ Post 132623 .