Recursive, an AI startup founded in 2025 and led by Richard Socher, Salesforce's former chief scientist, released Flint, its first model, on October 21, 2026. It is a mixture-of-experts (MoE) model aimed at agentic coding on local hardware: 40 billion parameters in total, with about 2.5 billion active per token. The weights are on Hugging Face under the Apache 2.0 license, and the context window goes up to 512k tokens, with a recommendation to stay within 256k for the best coding results.
The pitch is coding offline, without sending code to the cloud or paying for an API. The company says Flint exceeds 125 tokens per second on a MacBook Pro with an M4 Pro and 150 tokens per second on an NVIDIA RTX Spark, at short contexts. It works with llama.cpp, Ollama, LM Studio, vLLM and SGLang, among others. On benchmarks, Recursive claims Flint leads MoE models under 50 billion parameters on tests such as SWE-Bench Verified and Terminal-Bench, but those numbers come from the company and have not yet been independently verified.
The most interesting part is how it was made: according to Recursive, the Eureka Machine, its automated research system, took part in nearly every stage, from architecture search to quantization and optimized kernels, and a small team of people and agents built the model from scratch in under four months. The company itself recommends breaking complex tasks into smaller, clearly described pieces.
Original source: Recursive
Other references: