DownloadRecursive Flint is an open-weight Mixture-of-Experts language model developed by Recursive for autonomous coding agents and software engineering. It contains 39.8B total parameters with approximately 2.5B activated per token, combining 49 Transformer layers with 256 routed experts per MoE layer. Its hybrid attention architecture uses 35 sliding-window layers and 14 full-attention layers, supporting a 262K-token context window. Flint is designed to investigate repositories, fix bugs, implement features, write tests, and execute multi-step workflows through reasoning and tool calling. It achieved 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified in developer-reported evaluations. Trained from scratch using pretraining, supervised fine-tuning, and reinforcement learning, it supports local deployment on NVIDIA H200 or B300 GPUs through SGLang. The BF16 checkpoint is released under Apache 2.0.
Features
- 39.8B total parameters with approximately 2.5B active per token
- 49-layer decoder-only Mixture-of-Experts Transformer
- 256 routed experts with six selected plus one shared expert
- 262K-token context window for extended coding workflows
- Hybrid sliding-window and full-attention architecture
- Multi-step reasoning and OpenAI-compatible tool calling
- Strong performance on SWE-bench and Terminal-Bench evaluations
- BF16 weights with SGLang deployment on NVIDIA H200 and B300 GPUs
Project Samples
