AnalysisAI ModelsSeptember 30, 2026

Victoria and Maple: two open-weights fine-tunes of Qwen3.8-Flash-Next

Read original source →reddit.com

Victoria prunes Qwen3.8-Flash-Next by 44% using REAP (512 to 288 experts per layer), then retrains at 4-bit NVFP4, scoring 70% on Terminal-Bench 2.1 with GGUF weights included. Maple is a Canada-first fine-tune; both were built on a Dell B300 in the lab.

1 source

More stories today

Open the live feed