Recursive Releases Flint: 40B-Parameter Local Coding MoE Model

Recursive released Flint, a 40-billion-parameter mixture-of-experts coding model that activates 2.5 billion parameters per token and runs locally on laptops. Flint leads MoE models under 50 billion parameters on agentic coding benchmarks and generates over 125 tokens per second on an M4 Pro MacBook Pro. The model was built from scratch in under four months with the company's Eureka Machine automated research system.
Event version 1.
Key insights
Top agentic coding performance in class
Flint leads MoE models with fewer than 50 billion total parameters on agentic coding benchmarks including SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified. It rivals dense models with 25-40 billion parameters while running more than five times faster.
Fast local inference
Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts. It runs up to twice as fast as comparable models.
Built with automated research
Recursive's Eureka Machine automated the research loop of ideation, implementation, and validation and contributed to nearly every part of Flint's model-building stack. Its contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels.
Local privacy and open weights
Flint supports context windows up to 512k tokens, with a recommendation to stay within 256k tokens. It runs locally so data stays on the user's computer and works offline, with open weights available on Hugging Face and support for llama.cpp, Ollama, LM Studio, vLLM, SGLang, and Unsloth.