Home/Vibe coding/Flint

Flint

Recursive's open-weight coding model for local and offline use, with about 40 billion total parameters.

Flint is an open-weight mixture-of-experts coding model from Recursive. It has 39.8 billion total parameters (about 40 billion) and about 2.5 billion active per token. Apache 2.0 weights are on Hugging Face: RecursiveSI/flint (BF16, about 80 GB) and RecursiveSI/flint-int4 (MLX INT4, about 22.6 GB, for Apple Silicon). Recursive says it supports up to 512K context and recommends staying within about 256K. It works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang. Recursive also ships rec, its own inference engine for Apple Silicon (M4 Pro, M5 Pro, and M5 Max), which can launch pi or OpenCode. Lab-reported: 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1, and 125 or more tokens per second on an M4 Pro MacBook Pro at short contexts. Recursive says it may need more guidance than large closed models on complex tasks.

• Best open-weight coding models to run locally