Recursive Flint is a newly released open-weight coding model from Recursive Superintelligence. The company describes it as a 39.8-billion-parameter mixture-of-experts (MoE) model that activates about 2.5 billion parameters per token, aiming to give coding agents a smaller active compute footprint than a similarly sized dense model.
What is Recursive Flint?
Flint is a text-based model built for software-development workflows: examining repositories, fixing bugs, implementing features, writing tests and interacting with tools. Recursive has published the model weights, tokenizer and serving resources on its official Hugging Face repository under the Apache 2.0 license. The open-weight license applies to the checkpoint; anyone using a hosted service should separately check that service’s terms and data practices.
Flint’s 40B total and 2.5B active parameters explained
A mixture-of-experts network routes each token through only a selection of its available expert components. Flint’s documented architecture has 49 layers, including 47 MoE layers. Each MoE layer has 256 routed experts, with six selected for a token and an additional shared expert. This helps explain why the model’s roughly 40 billion total parameters are not all active for every token. It does not mean the full model occupies only 2.5 billion parameters of storage.
The published configuration specifies a 262,144-token context window, shared by the prompt and generated response. That figure is a capacity setting rather than a guarantee of equal accuracy throughout a very long context.
What do the benchmark results show?
Recursive reports 74.2% on SWE-bench Verified, 54.9% on Terminal-Bench 2.1 and 89.7% on SWE-bench Pro v2 for its BF16 checkpoint. These are developer-reported results, not independently reproduced measurements by TheNews92. Comparisons across model makers need caution because test harnesses, internet access, sampling settings and scoring rules can differ. Recursive also notes that its Terminal-Bench evaluation counted solutions obtained through web retrieval as unsuccessful.
Can developers run Flint locally?
The official model card provides an SGLang serving kit and describes a BF16 checkpoint. Its documented deployment example uses an NVIDIA H200 or B300 GPU, while quantized variants are prepared separately. Hardware, memory and software requirements depend on the checkpoint and serving configuration. Developers should consult the current repository instructions rather than assuming that any laptop can run the full BF16 model.
What should teams test before adopting it?
A practical evaluation should use a representative codebase and check patch correctness, test pass rates, tool-call safety, latency and memory consumption. Model-generated code still requires review, and automated agents should be restricted to appropriate file and command permissions. Recursive’s model card explicitly cautions that Flint can produce faulty code or inappropriate tool calls.
Bottom line
Flint adds another open-weight option for coding-agent experiments, with a notably low active-parameter count relative to its total size. Its suitability for production depends on independent tests on the actual workloads and hardware involved, not a single headline benchmark.
Source and disclosure: Technical specifications and benchmark figures come from Recursive’s official model card, accessed 22 October 2026. TheNews92 did not independently benchmark the model.
