StudioRecursive Releases Open-Source MoE Coding Model Flint (39.8B Total Parameters, ~2.5B Activated)
Original title: RecursiveSI/flint (text-generation)
Recursive has open-sourced the coding agent model Flint on Hugging Face, featuring a decoder-only MoE architecture with 39.8B total parameters and approximately 2.5B activated per token. Supporting a context window of 262,144 tokens under an Apache 2.0 license, Flint achieves 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. BF16 weights and an SGLang serving suite are available for deployment on a single B300 or H200 GPU; quantized versions for NVIDIA and Apple Silicon are forthcoming, while Transformers modeling code is not yet included.
The README provides complete architectural details, benchmark scores, and single-GPU deployment instructions, enabling readers to assess whether this open-source coding model fits their agent workflows.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.