Recursive releases Flint — a 39.8B-parameter MoE model for coding agents

On October 21, 2026, the model RecursiveSI/flint was published on HuggingFace with the text-generation task. It is a mixture-of-experts language model from Recursive aimed at coding agents: exploring codebases, fixing bugs, implementing features, and writing tests. It supports text input and output, reasoning, and multi-step tool use. In the author's evaluations, the model scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.

What it means

The full model weighs 74.1 GB in safetensors format, with the largest file at 5 GB. The license is apache-2.0, it runs via transformers, and the repository contains a BF16 checkpoint and an SGLang serving kit. The architecture is a decoder-only mixture-of-experts Transformer: 39.8 billion total parameters, about 2.5 billion active per token, 49 layers, 47 of them MoE. For local running, the site has 16 quantized builds — GGUF, MLX, NVFP4, GPTQ, and FP8. The smallest working build is 20.7 GB in IQ4_XS format from terelabs/flint-GGUF. The author notes that the quantized variants are prepared separately for NVIDIA and Apple Silicon. The model has just appeared, downloads are still at zero, but it already has 289 likes. You can follow new builds and local tools in the CPU3D news feed.