Recursive releases Flint coding model built largely by AI — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Recursive released Flint on Wednesday as a free, open-source coding model under the Apache 2.0 license on Hugging Face, designed to run locally so developers can code offline and avoid cloud usage fees.
- Flint is a mixture-of-experts model with 40B total parameters that activates only 2.5B per task, hitting ~131 tokens/sec on a MacBook Pro M4 Pro — roughly twice the speed of Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning on the same hardware.
- Recursive's Eureka Machine automated nearly every step of Flint's development — pre-training, post-training, and inference — including a low-bit training technique that shrank the model by 71% with minimal quality loss, letting a small team ship the model in under four months.
- Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from GV, Nvidia, and AMD Ventures, and is led by CEO Richard Socher, Salesforce's former chief scientist.
- Flint topped SWE-Bench Verified among models under 50B parameters, and on Recursive's average across six coding tests, only Alibaba's Qwen3.8 27B scored higher.
- Recursive told TechCrunch its $410 million compute deal with AWS in June was 'less about headcount and more about agent count,' framing its long-term goal as full recursive self-improvement with humans able to stop the process.
Why it matters: If Recursive's approach holds, each new model could be built faster and cheaper by AI agents rather than human teams, potentially reshaping the economics of model development. For developers today, Flint delivers a fast, free, offline-capable coding assistant that keeps proprietary code off third-party clouds — directly undercutting paid services that charge per token processed.
Ask SkimNews
