FULL STORY

dots3-note: 280B Open Agentic Model Goes Local

dots-studio released the dots3-note preview, a 280B open multimodal MoE model for long-horizon agentic tasks. It was quickly integrated into llama.cpp for local runs, scoring 78.4 on SWE-bench.

2026-08-16 ~ 2026-08-17 · 2 episodes · 5 posts

Episode 1 · llama.cpp adds Dots3 Note model with SWE-bench 78.4 score (2026-08-16, 2 posts)

llama.cpp integrated the dots3-note-prev model (280B/16B active), which achieved a 78.4 score on SWE-bench Verified. The model, released by dots-studio for local execution, features DSA (Dense Sparse Attention) and SW optimizations.

Episode 2 · dots3-note Preview: 280B Open Multimodal Model for Long-Horizon Agents (2026-08-17, 3 posts)

dots3-note preview is an open-weight 280B-parameter (16B active) multimodal MoE model for long-horizon agents, introducing TEMPO, a reinforcement learning method using self-critique and time-scaled value estimation for long-horizon training.