Recursive releases Flint, a free 40B coding model for laptops

Recursive has released Flint, its first model: a free coding model small enough to run on a laptop. It is a mixture-of-experts model with roughly 40B total parameters and about 2.5B active per token. The weights are on Hugging Face under the Apache 2.0 license. The company says its automated research system, called the Eureka Machine, helped build nearly every part of the stack.

Flint is built for coding agents: investigating codebases, fixing bugs, implementing features and writing tests. It is text-only and supports reasoning and OpenAI-compatible tool calls. The model card documents a 262,144-token context window. Recursive reports 74.2% on SWE-bench Verified and 95.1% on IFEval, but only 54.9% on Terminal-Bench 2.1, where it trails a dense Qwen model at 73.0%. Recursive's own chart puts it on top among mixture-of-experts models under 50B on agentic coding, while still behind the strongest dense models.

Hardware requirements are modest. The 4-bit weights take about 23 GB, and the model is described as running on 48 GB Apple Silicon Macs or 48 GB NVIDIA GPUs. Community builds and some listings say it fits in 24 GB. Recursive claims 125+ tokens per second on a MacBook Pro and 150+ on RTX Spark. Support is already broad: Ollama runs it on Mac for now, with other GPU support promised in a later release, and llama.cpp has added the model architecture. Dynamic GGUF and MLX builds have also appeared, and NVIDIA lists it for its local hardware. Recursive positions Flint as an on-device complement to frontier cloud agents, not a replacement.

Our take

A permissively licensed model that scores 74.2% on SWE-bench Verified while activating only 2.5B parameters is a real step for local coding agents. The numbers are Recursive's own, though, and the weak Terminal-Bench result suggests it is not a drop-in replacement for the best dense models.