Recursive has released Flint, a language model for programming tasks that the company says was developed largely with help from its internal system, Eureka Machine. The system proposed experiments, built and tested them, and used results to plan follow-up work. Flint is the first public model from that process, offering a concrete example of AI participating in AI research.
How Eureka Machine shaped Flint
Recursive says Eureka Machine worked across training data, optimizers, model architecture, and inference tuning, running multiple approaches in parallel. The company describes this as an automated research cycle, with the system contributing at several stages rather than handling only one narrow task.
One example involved testing expert configurations for a mixture-of-experts model. The work began with an architecture using 35 billion total parameters and 3 billion active parameters. Recursive says the selected design, with 40 billion total parameters and 2.5 billion active per token, offered the best balance of speed and coding performance among the variants tested.
The system also explored quantization, which reduces the precision used to represent model weights. Recursive says Eureka Machine applied four-bit quantization to the output layer, then added a distillation step intended to recover accuracy. The company reports that the resulting NVFP4 model is 0.7 points behind the 16-bit version across benchmarks and that the approach needed roughly half as many training steps as the previous method.
Built for coding on local hardware
Flint is designed for agentic coding and local execution. Its mixture-of-experts architecture has 40 billion parameters in total, but only 2.5 billion are active for each token. Recursive says that makes the model a candidate to compete with dense models in the 25 to 40 billion parameter range while running more than five times faster.
The company reports generation speeds of more than 125 tokens per second at short contexts on a MacBook Pro with M4 Pro, and more than 150 on an Nvidia RTX Spark with 64 GB of memory. It also says inference can be up to twice as fast as with similar models. Recursive claims Flint leads coding benchmarks including SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified among MoE models with fewer than 50 billion parameters. Those results have not been independently verified.
Hardware-specific work is part of the design. Recursive says Eureka Machine created compute kernels for Apple Silicon and Nvidia RTX Spark. On Apple's M4 and M5 chips, a Metal megakernel combines the forward pass during token generation into a single operation. The company attributes Flint's speed to optimizations across multiple parts of development.
Open weights, local use, and practical limits
Flint's weights are available on Hugging Face under the Apache 2.0 license. The model supports llama.cpp, Ollama, LM Studio, vLLM, and SGLang; Recursive also released rec, its own inference engine for Apple Silicon. The context window can reach 512,000 tokens, although Recursive recommends a limit of 256,000.
Local execution means code can be processed without sending it to external servers. Recursive also acknowledges a limitation: Flint needs more guidance on complex tasks than large proprietary coding models. Its recommendation is to divide bigger tasks into smaller steps, which gives users a practical way to work within that constraint.
A first result, not proof of self-improvement
Recursive was founded by Richard Socher and others, including former DeepMind researcher Tim Rocktäschel. The startup came out of stealth in May 2026 and raised $650 million at a $4.65 billion valuation. At that point, it had no concrete technical results to show. Flint is now its first publicly available model built through the automated research process.
The company's longer-term ambition is to use each new model to make Eureka Machine a better researcher, helping build later models faster and more efficiently. Recursive eventually wants to apply the same approach to research in other scientific fields. That is a much broader goal than producing a single coding model.
Flint alone does not establish that the loop works recursively. To demonstrate that, Recursive would need to show that a model developed with Eureka Machine makes the system itself better at research. Finding promising ideas and judging results reliably remain difficult for AI systems, and the degree to which AI agents can take over research work is still debated. Flint is evidence of AI-assisted model development; it is not yet evidence of a system that can reliably improve its own research capability.