Recursive's Flint: a new MoE model for coding locally

Recursive has announced Flint (TechCrunch reported the launch on October 21, 2026; the official page says October 23, 2026). The mixture-of-experts model is designed to run locally on devices such as MacBook Pros with Apple M4 Pro and M5 Pro/Max chips or the 64 GB NVIDIA RTX Spark. It has 40 billion total parameters (39.8 billion according to Hugging Face) and activates about 2.5 billion parameters per token. The weights are released under the Apache 2.0 license on Hugging Face (repo RecursiveSI/flint) and work with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. Alongside the model, the company shipped “rec”, a lightweight inference engine optimized for Apple M4 Pro and M5 Pro/Max chips.

According to Recursive, Flint reaches more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark with short contexts, speeds the company puts at up to twice those of comparable models. TechCrunch reports about 131 tokens/s on the MacBook, consistent with the claimed figure. The context window is given as 512K tokens in the announcement, but the Hugging Face model card lists 262,144 tokens.

The release benchmarks, published on the Hugging Face card, put Flint at 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. These are Recursive's own evaluations and have not yet been confirmed by independent testing. According to the same authors, the NVFP4 quantized version stays within 0.7 points of the 16-bit model. Recursive claims Flint is the best MoE model under 50 billion total parameters on agentic coding benchmarks, and says it rivals dense models of 25–40 billion parameters while running more than five times faster.

Recursive says Flint was built by "a small team of people and agents" in under four months, without saying how many people. According to the company, its automated research system, Eureka Machine, contributed to almost every part of the work: architecture, quantization, distillation, data, training and inference kernels. How much it really did compared with the humans can't be measured from the outside.

The open-weight race is shifting toward local execution. A week earlier, Meta had released Muse Spark Open (816 billion parameters, needing at least 1 TB of memory). Flint goes the opposite way: a small, sparse MoE for coding offline with no per-token costs.

— Pixie