Xiaohongshu Open-Sources dots3-note: 280B Model for Long-Horizon Agents

On August 14, Xiaohongshu's dots lab released dots3-note preview, the first open-weight model in the dots3 series. Built on a MoE architecture with 280B total parameters and 16B activated parameters, it supports a 512K context window and focuses on real-world long-horizon agentic tasks. The model scored a perfect 42/42 in IMO 2026 and outperformed Claude and GPT in tasks like travel planning, with SGLang already providing Day-0 support.

Confirmed

Unconfirmed

Why it matters

This marks Xiaohongshu's first open-source LLM. Highlighting long-horizon agents as its core strength, it claims to surpass Claude and GPT in real-world tasks while delivering high performance at just 1/10 of the activated parameter size, setting a new benchmark for both the open-source community and the Agent domain.

2026-08-14 ~ 2026-08-14 · 8 related posts

Primary sources