China's Instagram Ships Its First AI Model: dots3-note Scores 75.1 on Terminal-Bench 2.1 Under Apache 2.0
Xiaohongshu — the platform Chinese users call Little Red Book, best described as Instagram crossed with Pinterest — has released its first frontier-class language model. dots3-note Preview shipped on August 14, with weights immediately available on Hugging Face under Apache 2.0 and an API live on OpenRouter.
The model is a Mixture-of-Experts architecture: 280 billion total parameters, 16 billion activated per token, with 256 routed experts plus one shared expert and a top-8 routing scheme. Context window is 512K tokens. Input modalities cover text, images, video, and audio; output is text only. The model uses a 7B visual encoder (MoE, 1.2B active) and an 800M dense audio encoder.
Terminal-Bench 2.1 Scores 75.1
Xiaohongshu’s team ran dots3-note on Terminal-Bench 2.1 using the Terminus-2 harness, scoring 75.1. That places it ahead of HunYuan Hy3 (71.7%) and DeepSeek V4 Flash’s 61.8% build but behind GLM-5.2 (81.0%), Kimi K3 (88.3%), and the closed frontier cluster. SemiAnalysis, which released its own chart of the results, called the score 4.9 points above the best-performing US open-weight model on its comparison chart.
These are self-reported numbers run by the releasing lab. Independent re-runs on the same harness will determine whether the claim holds. The dots studio team published detailed evaluation notes: 8 CPUs and 16 GB RAM per task, 10-hour timeout, temperature 0.7, top-p 0.95, max 81,920 tokens.
What TEMPO Is
The training method behind the model’s agent performance is called TEMPO — the lab’s reinforcement learning approach for long-horizon agent tasks. A technical report is promised within a week of release. The model architecture layers a single dense layer plus 45 MoE layers, with an alternating pattern of 13 dense-sparse attention layers and 33 sliding-window attention layers.
The dots3 family is structured as three releases: note (the current, lightest), jazz, and aria. The release materials describe note as covering “lower levels of task complexity” with lower computational cost, suggesting jazz and aria will scale the architecture rather than simply change fine-tuning targets.
Why This Is Worth Watching
Most frontier open-weight models come from AI labs, infrastructure companies, or hyperscalers. Xiaohongshu is a consumer social platform with roughly 300 million monthly active users — primarily in China, primarily for lifestyle content. Its AI lab, dots.studio, built this model without any of the market-side AI positioning that drove comparable releases from Alibaba, ByteDance, or Tencent.
The Apache 2.0 license is clean commercial use with no revenue caps, unlike the Kimi K3 licence (which carries revenue triggers) or MiniMax H3 (capped at $20M revenue). For teams that want to build on or fine-tune a large open MoE without licence complications, that distinction matters.
The SWE-bench Verified number is not yet available. For teams selecting coding agents, dots3-note’s Terminal-Bench performance is a strong signal but the agentic coding harness result is still pending. The technical report and independent evaluations in the coming week will either confirm or complicate the launch benchmarks.