vLLM Adds Day-0 Support for Recursive's Flint Model

Recursive AI has released Flint, a 40B Mixture-of-Experts coding model built by its automated research system, Eureka Machine, with vLLM providing immediate Day-0 support. This launch sets new benchmarks for both local and cloud inference efficiency while maintaining an open-source Apache 2.0 license.

Technically, Flint activates only ~2.5B parameters per token and supports a 262K context window. On consumer hardware, it achieves over 125 tokens/s on a 48GB MacBook Pro—7-10x faster than comparable models like Qwen3.8. In data centers, optimized kernels in vLLM enable 500 tokens/s decoding on a single GB200 GPU using NVFP4 or BF16 quantization.

This development highlights the industrial viability of AI-generated models. By balancing high code quality (~81% average benchmark score) with extreme speed, Flint significantly lowers the barrier for developers to deploy high-performance, private coding assistants via ready-to-use Docker images.