Flint: A Small MoE Model

RecursiveSI released Flint, a 39B-parameter mixture-of-experts model with 2.5B active parameters per token, built from a 2-dense-layer, 47-MoE-layer architecture and trained from scratch through tokenizer, pretrain, midtrain, SFT and RL stages. Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, with the team reporting the full training pipeline was built in 4 months. Training from scratch gives more flexibility for downstream fine-tuning.