Flint is a 40B parameter open weight machine learning system that achieves speeds of over 125 tokens per second on M4 Pro MacBook Pro computers. The mixture-of-experts model uses 2.5B active parameters per token and leads competitors under 50B total parameters on agentic benchmarks. It runs at 150 tokens per second on Nvidia RTX Spark hardware and was developed from scratch in under four months by the Eureka Machine, an automated research system that designed the training recipes and inference kernels.
The model is released under an Apache 2.0 license and is available on Hugging Face, OpenCode, and Pidotdev. Collaborative optimization with vLLM reduced kernel launches from 22 to 7 per layer, enabling 500 tokens per second on a single Nvidia GB200 with 100K context. Josh Tobin of Recursive believes that a shift toward local execution on laptops and nearby servers will eventually define everyday software development to protect data privacy.