Recursive, a startup founded last year with the goal of building AI that can improve itself, has released its first model: Flint. It is a coding model with 40 billion total parameters, of which only 2.5 billion are active for each token, designed to run locally on your own computer. The weights are available on Hugging Face under the Apache 2.0 license, so anyone can download them, run them offline and adapt them to their own applications.
The most unusual aspect concerns how the model was built. According to the company, a small team of people and agents built it from scratch in under four months, with the help of an automated research system called the Eureka Machine. The system proposes ideas, implements them, tests them and uses the results to decide what to try next, exploring several directions in parallel. Recursive says it contributed to nearly every part of development, from data and training recipes to infrastructure and inference optimization.
Image: recursive.com
Performance on coding benchmarks
According to the company's data, Flint is the best-performing MoE model with fewer than 50 billion total parameters on agentic coding benchmarks, including SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. Recursive also claims it rivals dense models with 25 to 40 billion parameters in coding while running more than five times faster. The results reported by the company are 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, obtained in its own evaluations.
The claimed speed is more than 125 tokens per second on a MacBook Pro with an M4 Pro chip and more than 150 tokens per second on a 64 GB NVIDIA machine, at short contexts. Again, these figures come from the company. The model supports context windows of up to 512,000 tokens, although Recursive recommends staying within 256,000 for the best results.
Where the model falls short
Recursive warns that Flint, while very capable, may need more detailed instructions than the large closed models developers are used to. For complex tasks, the suggestion is to split the work into smaller subtasks.
Flint + Flint | Haircare packaging, Skin secrets, Beauty skin care
How the model was built
According to the company's documentation, the architecture was chosen by the automated system, which explored different expert configurations starting from a base of 35 billion total and 3 billion active parameters. After training, the same system searched for quantization recipes to reduce memory requirements: the output layer was quantized to 4 bits, against common practice, and accuracy was recovered through a distillation stage. The final NVFP4 model is within 0.7 points of the 16-bit version on the company's benchmarks.
Recursive has also released a lightweight inference engine optimized for Flint on M4 Pro and M5 chips. Flint works with tools such as llama.cpp, Ollama, LM Studio, vLLM and SGLang.
The company's stated goal is to build a research process that improves with each new model it produces. Flint, according to Recursive, is the starting point of that path.
No comments yet. Be the first!