Recursive unveils Flint, a high-performance 40b moe coding model for AI agents

Recursive Superintelligence has launched Flint, a new open-weight 40B MoE model designed for coding agents. The model, built with the Eureka Machine, features 2.5B active parameters and supports reasoning and OpenAI-compatible tool calls. Flint is a decoder-only MoE Transformer capable of investigating codebases, fixing bugs, implementing features, and writing tests. It operates on Hugging Face under the Apache 2.0 license. Flint's architecture includes 49 layers, with 2 dense layers followed by 47 MoE layers. Each MoE layer routes 6 experts per token, plus a shared expert. The model achieves speeds of over 125 tokens/sec on an M4 Pro MacBook Pro and 150 tokens/sec on an RTX Spark, claiming up to 2x faster performance than comparable models. Recursive highlights its efficiency, stating it runs more than 5x faster than dense 25-40B models with similar coding scores. The model supports multiple local options, including integration with llama.cpp, Ollama, and vLLM. Recursive reports scores on benchmarks like SWE-bench and Terminal-Bench, though it notes that Qwen3.6 outperforms Flint in some metrics. The company emphasizes its contributions to the field, with technical details and model weights available on GitHub.