FULL STORY

dots3-note: Open-Source Debut and Real-World Agent Test

dots studio open-sourced dots3-note Preview, a 280B tri-modal agent model, on August 29. Developers soon tested it by building a wedding-coordination agent, demonstrating its long-horizon planning capabilities.

2026-08-29 ~ 2026-08-31 · 2 episodes · 11 posts

Episode 1 · Rednote Open-Sources dots3-note: 280B Multimodal Agent Model with 512K Context (2026-08-29, 9 posts)

On August 29, Rednote's AI lab dots studio released dots3-note Preview, an open-source multimodal agent model built for long-horizon, evolving real-world tasks. It has 280B total / 16B active parameters (MoE), a 512K context window, text-vision-audio understanding, and ships under Apache 2.0 — the first tri-modal model in the dots3 family — with early benchmark results drawing community attention.

Confirmed

  • Specs: 280B total / 16B active parameters, 512K context, Apache 2.0 license
  • Multimodal: joint understanding of text, vision, and audio; first tri-modal model in the dots3 family
  • Positioning: workflows requiring sustained reasoning, execution, and dynamic adaptation, addressing plan coherence in long-horizon agent tasks
  • Training: per @heyshrutimishra, the core innovation is the TEMPO training method targeting task persistence
  • Capability loop: stronger coding, tool calling, and end-to-end delivery close the loop from perception to action
  • Official demo: autonomously building a Xiaohongshu (Rednote) VR experience
  • Benchmarks: per @heyshrutimishra, ARC-AGI-2 score of 81.39 and SWE-bench Verified of 78.4%

Unconfirmed

  • Exact evaluation settings and reproducibility of the benchmark scores await official technical reports or independent community testing

Why it matters

dots3-note is among the few open models combining ultra-long context with tri-modal understanding under a permissive license, offering a new foundation for agent applications needing long-term memory and multimodal interaction. Its 512K context and sparse activation design target long-horizon coherence, and its reported scores on hard benchmarks like ARC-AGI-2 and SWE-bench Verified make community replication worth watching.

Episode 2 · dots3-note Demonstrated as Long-Horizon Agent in Wedding Planning Test (2026-08-30, 2 posts)

A developer tested dots3-note Preview as a wedding-coordination agent handling 120 guests, a $35k budget, and disruptions like storms. A technical blog and free API were released so readers can reproduce the experiment.