Recursive Flint: a 40B open coding model for laptops that uses 2.5B parameters per token
What is it?
Flint is the first model from Recursive, the company Richard Socher founded. It's a mixture-of-experts (MoE) model with 40B total parameters, of which only 2.5B are active per token, so it runs fast on a laptop. Context goes up to 512k tokens, and the company recommends staying at 256k or below. The weights are on Hugging Face under Apache 2.0 and work with llama.cpp, Ollama, LM Studio, vLLM and SGLang.
hf download RecursiveSI/flintWhy does it matter?
Recursive reports more than 125 tokens per second on an M4 Pro MacBook ProVendor claim and more than 150 on an RTX SparkVendor claim. It also says Flint leads MoE models under 50B on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench VerifiedVendor claim, and that the NVFP4 quantized model lands within 0.7 points of the 16-bit oneVendor claim. Teams that can't send code to an outside API can try it as a local coding agent. The company says much of the model was built by its own automated research system, the Eureka Machine.
Maturity
GAThe weights are out under Apache 2.0, so you can download and use it today. The launch post gives no numeric benchmark scores.
How to track it
- Independent tests: watch for Flint scores on Artificial Analysis or the official SWE-bench leaderboard.
- Local runtimes: check whether Flint shows up in the official Ollama and LM Studio model lists.
- What's next: follow the Recursive blog and Richard Socher on X for a bigger Flint or a next model.