Recursive Releases Flint, a 40B Open Coding Model
Recursive has released Flint under Apache 2.0, pairing a sparse 40B architecture with local deployment for coding agents.
Recursive has released Flint, an open-weight coding model aimed at agents that investigate repositories, fix bugs, implement features and write tests. The model is available through Hugging Face under the commercially permissive Apache 2.0 license.
Flint has 39.8 billion total parameters but activates approximately 2.5 billion per token through a mixture-of-experts design. Its model card lists a 262,000-token context window and support for reasoning, multi-turn tool use and OpenAI-compatible tool calls. Recursive says the model can run on a single B300 or H200 GPU, while quantized variants target NVIDIA and Apple Silicon runtimes.
The release is notable because it targets the gap between large hosted coding models and smaller local systems. Running an agent repeatedly requires more than benchmark quality: teams need predictable latency, manageable infrastructure and control over source code and prompts. Sparse activation reduces the computation used for each token, while the Apache license removes many of the commercial restrictions associated with proprietary services.
Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. These figures are company-reported and use particular agents, sampling settings and evaluation procedures, so they should not be read as a universal ranking. The model’s roughly 80GB BF16 checkpoint also means that “local” does not necessarily mean laptop-friendly without quantization or specialized hardware.
The surrounding ecosystem moved quickly. vLLM and OpenHands announced day-one support, making Flint usable inside established coding-agent workflows rather than leaving developers with an isolated checkpoint. That lowers adoption friction and gives the release a clearer path into experimentation and production trials.
Flint will matter if its sparse design delivers reliable repository-level work at a lower recurring cost, not merely if it posts strong benchmark numbers. Its strongest immediate contribution is widening the set of commercially usable models that organizations can host and adapt themselves.
Uncle Cat take
Flint’s Apache 2.0 license is more consequential than its headline scores, but the 80GB BF16 download keeps “local” deployment a specialist proposition.