Recursive's Flint: a 40B open coding agent you can run locally
Apache-2.0 weights, 74% on SWE-bench Verified, 125+ tok/s on an M4 Pro, and free on OpenRouter — plus a new tool for which model fits your GPU.
Copy markdown
A laptop-class coding model with near-frontier scores
Flint is a 39.8B-parameter mixture-of-experts model with just ~2.5B active per token, built for agents that investigate codebases, fix bugs and write tests, with a 262K-token context window. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1 — numbers that put a sub-3B-active open model in the same bracket as hosted frontier coders.
Run it tonight — on a Mac, or free on OpenRouter
The 4-bit weights are ~23GB and fit a 48GB Apple Silicon Mac (M4/M5 Pro), hitting 125+ tokens/sec on an M4 Pro. It ships in GGUF, NVFP4, MLX and BF16 and runs under llama.cpp, Ollama, LM Studio, vLLM and SGLang; there's also a free 262K-context OpenRouter endpoint and Recursive's own "rec" engine for Apple Silicon. Apache 2.0, so you can wire it straight into your own agent.
The pitch: an AI built most of it
Recursive says its automated "Eureka Machine" drove nearly the whole stack — architecture search, quantization, kernels — and shipped Flint in under four months. Take the self-reported benchmarks with caution (its SWE-bench Pro figure isn't comparable to the public leaderboard), but the weights are open, so you can test its real coding behavior yourself.
New tool: which local model actually fits your GPU
quantized.uk, posted to Hacker News this morning, indexes 117 models across GGUF/AWQ/EXL2/GPTQ and 78 GPUs, then tells you which quant fits your card, how much VRAM is left and the exact command to run it. The sizing is calculated, not measured, so double-check before a big pull — but it's a quick first filter for running something like Flint locally.