AI startup Recursive released Flint on Wednesday, a free coding model that runs directly on laptops without cloud connectivity or per-use fees. The move signals a deliberate challenge to the cloud-dependent AI business model that powers much of the sector's revenue today.
Founded last year, Recursive built Flint using its own AI research system, the Eureka Machine, in less than four months with a small team. The model implements recursive self-improvement—AI building and refining subsequent AI—a capability that historically required far larger engineering teams.
The company emerged from stealth in May with $650 million in Series A funding and a $4.65 billion valuation. Backers include GV, Nvidia, and AMD Ventures. Richard Socher, Salesforce's former chief scientist, serves as CEO.
Architecturally, Flint is a mixture-of-experts model that activates only 2.5 billion of its 40 billion parameters per task, allowing it to run on standard laptops. That efficiency undercuts Meta's Muse Spark Open, released last week for Macs, which requires 816 billion parameters and at least one terabyte of memory.
Performance tests show Flint generating approximately 131 tokens per second on an Apple MacBook Pro with an M4 Pro chip—roughly double the speed of Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning (58-69 tokens per second on the same hardware). On Nvidia's RTX Spark, Flint exceeded 150 tokens per second.
On the SWE-Bench Verified benchmark, which measures real-world bug-fixing ability, Flint outperformed comparable models under 50 billion parameters. Across six internal benchmarks, only Alibaba's Qwen3.8 27B model scored higher.
Flint is available free on Hugging Face under the Apache 2.0 open-source license. Recursive notes that Flint requires more explicit guidance than larger paid coding assistants and recommends decomposing complex tasks into smaller steps for best results.
The economics here matter more than the benchmark numbers. By eliminating cloud inference, Recursive sidesteps the per-API-call revenue model that justifies massive cloud AI infrastructure spending. That forces a question for investors in cloud-dependent AI startups: if a bootstrapped model—trained by the company's own AI system—can run circles around cloud alternatives on consumer hardware, what justifies expensive usage-based pricing. Recursive's path suggests the moat in coding AI shifts from compute access to algorithmic efficiency. That's a bet on silicon economics, not cloud lock-in.
