Recursive Flint: 40B MoE with 2.5B active for local coding agents, Apache 2.0
In one sentence On October 21 Recursive published Flint on Hugging Face, a MoE with 39.8B total and about 2.5B active parameters per token, a 262K-token context and an Apache 2.0 license. It claims 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, figures not yet independently verified.
Recursive (Recursive Superintelligence) released Flint, a model designed to run coding agents directly on a laptop instead of in the cloud. It has about 40 billion parameters but uses only 2.5 billion for each word, which makes it fast even on consumer hardware.
The license is Apache 2.0 with no usage thresholds: it can be downloaded, modified and used commercially. Compressed versions are already available for the most common local tools, such as llama.cpp, Ollama, LM Studio and Jan, plus a dedicated engine for Apple silicon Macs.
The claimed coding results are high for a model this size, but the company measured them itself. On other tests, such as Terminal-Bench, larger dense models stay ahead. Recursive says it built the model from scratch, helped by its own automated research system.
Companies
Recursive
Tools
Flint
Tags
Sources