REAP

Open-weights releases: Victoria and Maple fine-tunes

Fine-tunes of Qwen3.8-Flash-Next: Victoria uses REAP to cut 44% of experts, trained in NVFP4, scores 70.0% on Terminal-Bench 2.1 and 159/164 on HumanEval, runs at 280 tok/s on one B300 with 35% fewer output tokens. Maple focuses on Canadian questions, improving official source citation from 6.0% to 62.9% and fully correct answers from 6.6% to 21.8% on 600 held-out questions.