Xiaohongshu

Aug 15, 2026

Xiaohongshu open-sources dots3-note preview, posts 75.1 on Terminal-Bench 2.1

Xiaohongshu AI Lab, dots.studio, has open-sourced the dots3-note preview model, according to PANews. The model uses a 280-billion-parameter Mixture-of-Experts architecture with 16 billion activated parameters and offers a 512,000-token context window. It also supports multimodal understanding across text, vision, and audio, while introducing a reinforcement learning method called TEMPO for long-horizon agent training. Model weights are now available on Hugging Face, and the API has been integrated with OpenRouter. A chart shared by SemiAnalysis shows dots3-note scoring 75.1 on Terminal-Bench 2.1, which is 4.9 points higher than the best-performing U.S. open-weight model shown in that chart. At the same time, SemiAnalysis said its team is still testing the model in everyday use to judge real-world quality and check whether it may have been overly optimized for benchmark performance.

1060

Xiaohongshu

Aug 14, 2026

Xiaohongshu open-sources dots3-note preview, its first public dots3 model for long-horizon agent tasks

Xiaohongshu’s dots model lab has released dots3-note preview as the first open-source entry in its dots3 family, extending public access to the same model line that previously drew attention after dots-note-3.0 achieved an officially certified perfect 42 at the International Mathematical Olympiad. The newly released model has 280B total parameters, 16B active parameters, a 512K context window, and multimodal understanding across text, vision and speech. Rather than centering the release on math alone, Xiaohongshu positions dots3-note preview around long-horizon, open-ended tasks such as travel planning, wedding preparation, store operations and home design, where goals are fluid and answers are rarely unique. The company also outlined the training and evaluation stack behind the release. That includes TEMPO, short for Test-time-scaled Value Estimation with Macro-step Policy Optimization, a reinforcement learning method built for long trajectories; a Self-Critiquing mechanism used both in training and inference; and two newly open-sourced benchmarks, VibeSearchBench and VibeLifeBench. Xiaohongshu said the preview model can handle tasks ranging from playing Slay the Spire II and solving ARC-AGI 3 to reading floor plans, generating decoration proposals, and building a native Apple Vision Pro app end to end. According to the company, a full dots3-note release is planned in the near term, while the broader dots3 lineup will also include jazz and aria.

6880

Bit.Fan