Recursive, the AI startup that came out of stealth in May with $650 million in funding, released Flint on October 21, its first model and an open-weight coding model built to run locally. Flint is a mixture-of-experts model with 39.8 billion total parameters and roughly 2.5 billion active per token, and its weights are available on Hugging Face under an Apache 2.0 license.
Speed is the pitch. Recursive says Flint generates more than 125 tokens per second on a MacBook Pro with an M4 Pro chip and more than 150 tokens per second on Nvidia's RTX Spark. Co-founder Richard Socher has put that at roughly twice the speed of comparably sized models. The documented context window is 262,144 tokens, prompt and output combined.
On Recursive's release evaluations, Flint scores 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1. The model card says agentic scores are average pass@1 over at least five attempts, and that Terminal-Bench solutions found through web retrieval were counted as failures. Recursive says those results lead Flint's size class on all three benchmarks.
Key facts
- 01Recursive released Flint on October 21, its first model, with open weights on Hugging Face under Apache 2.0.
- 02Flint has 39.8B total parameters and about 2.5B active per token, with a 262,144-token context window.
- 03Recursive reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1.
- 04The company says Flint runs at more than 125 tokens per second on a MacBook Pro with an M4 Pro chip.
- 05Recursive emerged from stealth in May with $650M in funding at a $4.65B valuation.
Under the hood, Flint is a decoder-only Transformer with 49 layers: two dense layers followed by 47 mixture-of-experts layers. Each token is routed to 6 of 256 experts plus an always-active shared expert. Thirty-five layers use sliding-window attention with a 513-token window and 14 use full attention. Recursive trained the model from scratch through pretraining, mid-training, supervised fine-tuning and reinforcement learning.
For server deployment, Recursive says one B300 or H200 GPU per replica is enough using the SGLang serving kit included with the weights. Flint supports OpenAI-compatible tool calls, and the model card recommends budgeting at least 32,768 output tokens for hard problems. On the consumer side, the model is described as workable for agentic coding on machines with 48GB of RAM.
The more unusual claim is how Flint was made. Recursive says a small team built it from scratch in under four months, leaning heavily on the Eureka Machine, the company's automated research system. According to Socher, the system generated and tested ideas for nearly every stage of the pipeline, including the training method, the architecture and chip-specific code, while the human team mostly monitored the process and picked the best results.
Recursive was founded by Richard Socher, Tim Rocktäschel, Jeff Clune, Josh Tobin and Tim Shi, whose backgrounds span DeepMind, OpenAI, Uber AI and Salesforce. Its $650 million round valued the company at $4.65 billion, with backing from GV, Greycroft and the venture arms of Nvidia and AMD. The company's stated focus is recursive self-improvement, in which AI systems help design and improve later AI systems.
In an early independent hands-on test on a Mac with 48GB of memory, Flint completed three tasks: a command-line to-do list, a web-search research tool and a simple explanatory website. It ran fast throughout, though three tasks is a small sample rather than an audit.
The benchmark figures come from Recursive and have not yet been verified outside the company. Recursive has also not shown outsiders evidence that the Eureka Machine improves itself recursively, as opposed to helping build a model. The model card itself warns that Flint can produce faulty code or inappropriate tool calls, and that long reasoning can raise latency and consume the output budget.
Flint's weights ship as BF16 safetensors in a single checkpoint, with a stated knowledge cutoff of June 2026. Recursive says terms for its hosted API are separate from the Apache 2.0 license that covers the open weights.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
