Recursive says AI built most of its new Flint coding model

Recursive released Flint on October 21, a 40-billion-parameter mixture-of-experts coding model with 2.5 billion parameters active at a time, which CEO Richard Socher told Wired was developed largely by the startup's automated research system, the Eureka Machine, with a small human team mostly picking the best results. Recursive says Flint, shipped under an Apache 2.0 license with a 262,000-token context window, leads its size class on SWE-Bench Pro, SWE-Bench Verified and Terminal-Bench and runs at over 125 tokens per second on a MacBook Pro with an M4 Pro chip and more than 150 on Nvidia's RTX Spark, but those benchmark figures come from Recursive itself and have not been independently verified. Recursive, which emerged from stealth in May with $650 million in funding at a $4.65 billion valuation backed by GV, Greycroft and venture arms of Nvidia and AMD, has not shown outside proof that the Eureka Machine is measurably improving at discovering new training methods and architectures, the actual definition of recursive self-improvement.

Nearly every part of Flint's pipeline, training method, architecture, even chip-tuned code, came from a system Richard Socher's startup built to generate and test its own research ideas. The humans at Recursive mostly watched and picked winners. A small human team mostly watched and picked the best results, according to Richard Socher. His startup Recursive released Flint on October 21, an open-weight coding model the company says its automated research system, the Eureka Machine, developed largely on its own. In a Wired feature by Will Knight, Socher told the outlet that nearly every stage of Flint's pipeline, the training method, the model architecture, even code tuned for specific chips, came from ideas the Eureka Machine generated and tested itself. That's the whole bet behind Recursive, which came out of stealth in May with $650 million in funding at a $4.65 billion valuation, backed by GV, Greycroft, and venture arms of Nvidia and AMD. The company, founded by Socher alongside Tim Rocktäschel, Jeff Clune, Josh Tobin, and Tim Shi, all veterans of DeepMind, OpenAI, Uber AI, and Salesforce, isn't trying to build another chatbot. It wants to build an AI system that can find its own weaknesses, write its own benchmarks, and rewrite its own code to get better, without much human hand-holding. Flint is the first public proof that process produces something anyone can use. On paper, Flint looks solid. It's a 40-billion-parameter mixture-of-experts model with only 2.5 billion parameters active at a time, which is why it runs on a laptop instead of a data center. Socher says it hits over 125 tokens per second on a MacBook Pro with an M4 Pro chip and more than 150 on Nvidia's RTX Spark, roughly twice the speed of comparably sized models. It ships under an Apache 2.0 license with a 262,000-token context window, and Recursive says it leads its size class on SWE-Bench Pro, SWE-Bench Verified, and Terminal-Bench, three of the standard tests for how well a model handles real coding tasks. Knight's own testing was limited. He ran Flint on a Mac with 48 gigabytes of memory and gave it three jobs: building a command-line to-do list, a web-search research tool, and a simple explanatory website. It was fast, and it got the jobs done. That's a real result, but it's three small tasks, not an audit. And Wired was careful to flag the obvious catch: the benchmark numbers showing Flint ahead of its peers came from Recursive itself, not from an independent lab. Nobody outside the company has verified them yet. ForgeCode claims it beats Claude Code and Codex, but the fine print tells another story https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ Tailcall's ForgeCode tops its own TermBench 2.0 leaderboard at 81.8%, but on the independently run SWE-bench Verified it edges Codex 1 by just 0.6 points, 72.7% to 72.1%, a far narrower gap than its marketing suggests. - how ForgeCode compares to Claude Code performance https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ - ForgeCode TermBench 2.0 leaderboard score explained https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ There's a bigger distinction buried in that caveat, and it's the one that matters for anyone trying to judge how close recursive self-improvement actually is to working. A model that codes well is not the same thing as a system that keeps making itself smarter on its own. Flint is a snapshot, a finished product the Eureka Machine happened to help produce. Whether the Eureka Machine itself is getting measurably better at discovering new training methods, architectures, and optimizations each time it runs, the actual definition of recursive self-improvement, is a claim Recursive hasn't shown outsiders proof of yet. Socher's story to Wired is about process, not a demonstrated improvement curve. This lands around the same time as a new profile in which Demis Hassabis, chair of Google DeepMind, told WSJ Magazine he now expects artificial general intelligence within two to three years, pulling in a timeline he'd previously put closer to five. Investors chasing the AI capability race will want to read Flint as a data point in that direction. It isn't quite that, at least not yet. A fast, well-built open-weight coding model is a genuinely useful shipping product. A verified, independent benchmark showing an AI research system outperforming its own prior generation on novel problems is the actual evidence recursive self-improvement is real. Recursive has given the world the first. It hasn't given anyone the second. None of that makes Flint a non-event. Apache 2.0 licensing means any developer or company can run it, fine-tune it, or build a product on top of it today, free of the usage restrictions that come with most frontier models. For founders watching where coding-agent infrastructure is heading, a model that runs locally at competitive speed, instead of routing every request through a hosted API, changes the economics of building on it. That's worth paying attention to regardless of what it does or doesn't prove about superintelligence. Frankly, the gap between "our AI helped build this" and "our AI is recursively improving itself" is exactly the gap every lab chasing this story needs the next twelve months to close, with numbers nobody but them generated. Also read: Demis Hassabis now says AGI could arrive in just two to three years https://startupfortune.com/demis-hassabis-now-says-agi-could-arrive-in-just-two-to-three-years/ • Gartner says AI infrastructure spending will outrun revenue by $4.9 trillion https://startupfortune.com/gartner-says-ai-infrastructure-spending-will-outrun-revenue-by-49-trillion/ • ChainOpera AI token COAI jumps 30% as AI-crypto trade heats up again https://startupfortune.com/chainopera-ai-token-coai-jumps-30-as-ai-crypto-trade-heats-up-again/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.