AI startup Recursive released its first public model, Flint, on Wednesday. The company describes Flint as a free coding model that can run on users’ personal computers and says much of its development was assisted by its automated research system, the Eureka Machine.
What Flint is and how it was built
Flint is a mixture-of-experts model: out of its total 40 billion parameters, only a small slice — 2.5 billion parameters — is activated for any given task. That design keeps the model light enough to run on an ordinary laptop while retaining a large overall parameter count.
Recursive says the Eureka Machine played a role at nearly every step of Flint’s creation. The automated system generated ideas, executed experiments, and used results to guide subsequent decisions; with this workflow, a small team reportedly assembled the model from scratch in under four months.
Performance and practical details
In Recursive’s internal tests, Flint is fast: on a MacBook Pro with Apple’s M4 Pro chip it produced about 131 tokens per second, roughly double the 58–69 tokens per second range reported for similar models on the same machine (including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning). On Nvidia’s RTX Spark hardware, Flint exceeded 150 tokens per second.
On SWE-Bench Verified — a benchmark that measures real-world software bug fixes — Flint outperformed other models under 50 billion parameters. Across Recursive’s six-test average, only Alibaba’s Qwen3.8 27B scored higher than Flint.
Availability and licensing
Because Flint runs locally on a user’s machine rather than on cloud servers, developers can work offline, keep code private, and avoid cloud usage fees that are often charged per token processed. Recursive has made Flint available for free on Hugging Face under the Apache 2.0 open-source license, which permits free use and modification.
Technical contributions from the Eureka Machine
Recursive credits the Eureka Machine with several technical optimizations: a sparser model shape, low-bit training that reduced Flint’s size by 71% with minimal quality loss, and custom kernels to accelerate core computations. The company did not quantify how much researcher time Eureka saved, but framed these contributions as part of a strategy where each new model should make the system a more capable researcher.
Company background and strategy
Recursive emerged from stealth in May with $650 million in funding and a $4.65 billion valuation, backed by investors including GV, Nvidia, and AMD Ventures. The startup is led by CEO Richard Socher, former chief scientist at Salesforce. In June, Recursive disclosed a $410 million compute agreement with Amazon Web Services, which it described as focused more on agent capacity than on headcount.
Limits and early reviews
The company acknowledged Flint may need more guidance than larger paid coding assistants and recommended breaking large tasks into smaller steps. Early hands-on reviews were positive: Wired’s Will Knight used Flint on a Mac with 48 GB of memory to build several small applications (a to-do app and a web-research tool) and reported that Flint consistently satisfied his coding requests.
Why this matters
Flint is an early, tangible example of Recursive’s approach to recursive self-improvement — the idea that AI systems can help design and build improved AI models under human oversight. The release highlights potential advantages (faster iteration, smaller locally runnable models, reduced cloud costs) and contributes to ongoing industry discussions about the governance and safety of increasingly autonomous AI research systems.
Overall, Flint represents a noteworthy step in testing whether automated research systems can materially accelerate model development while remaining subject to human control.
