Xiaohongshu Open-Sources dots3-note: 280B MoE With 16B Active, 512K Context, 75.1 on Terminal-Bench 2.1

PANews·medium signalXBlueskyLinkedInCopy link

Xiaohongshu's dots.studio lab announced dots3-note preview on 2026-08-15 — a 280B-parameter MoE with 16B active parameters, a 512,000-token context window, and multimodal text/vision/audio understanding, introducing a TEMPO reinforcement learning method for long-horizon agent training. A SemiAnalysis chart puts it at 75.1 on Terminal-Bench 2.1, 4.9 points above the best US open-weight model shown. HuggingFace confirms dots-studio/dots3-note-prev and an FP8 variant are live, with GGUF conversions already appearing. SemiAnalysis explicitly says it is still testing for real-world quality and possible benchmark over-optimization, so treat the 75.1 as unconfirmed.