Xiaohongshu Open-Sources dots3-note 280B Multimodal Model

Xiaohongshu's Dots Lab open-sourced the first model of the dots3 series, dots3-note Preview. Featuring a MoE architecture with 280B total parameters and 16B active parameters, it supports 512K context length and native understanding of text, images, video, and audio. According to official sources and multiple reports, the model achieved a perfect score of 42/42 in IMO 2026 and outperformed Claude and GPT in long-horizon agent tasks such as travel planning and wedding preparations, with overall performance comparable to DeepSeek-V3.

Confirmed

• Release & Specs: @小红书技术REDtech confirmed the open-sourcing of dots3-notePreview, the lightest model in the series. It adopts a MoE architecture with 280B total parameters and 16B active parameters, supporting 512K context and multimodal understanding. The model is licensed under Apache 2.0.
• Performance: According to @机器之心 and @xiaosun86, the series achieved a perfect score of 42/42 in IMO 2026; the open-source version outperformed Claude and GPT in open-ended tasks like travel and wedding planning. @ChengleiSi reported that its overall performance surpasses HY3, matches or exceeds the GLM series, and rivals DeepSeek-V3.
• Tech & Ecosystem: The team proposed the TEMPO training paradigm, optimizing performance through value estimation and macro-strategies during test-time scaling. @jacek2023 and @xiaosun86 noted that SGLang has provided Day-0 support.
• Leak Incident: @teortaxesTex mentioned that the model architecture was previously leaked and mistakenly rumored to be ByteDance's Doubao model.

Unconfirmed

• Specific ARC-AGI benchmark scores and task settings were not disclosed in the materials; @teortaxesTex only mentioned that the scores were astonishing without providing specific numbers.

Why it matters

• This marks Xiaohongshu's first open-source large model, achieving performance comparable to top proprietary models with about 1/10th of the active parameters, providing a scarce option for the open-source community targeting long-horizon agent tasks without standard answers.

2026-08-14 ~ 2026-08-15 · 14 related posts

• [source] dots3-note Multimodal Model Released: 280B Params, 512K Context — jacek2023 · 2026-08-14
• Leaked ByteDance Doubao Dots3-note Specs: 280B Total Params, 512K Context — teortaxesTex · 2026-08-14
• Dots3-note Multimodal Model Leaked: 280B Params, 512K Context, Shocking ARC-AGI Scores — teortaxesTex · 2026-08-14
• Rednote's Dots Studio Releases First Open-Source LLM dots3-note: 280B MoE for Long-Horizon Agents — BanghuaZ · 2026-08-14
• [source] Xiaohongshu Open-Sources 280B Agent Model, Outperforming Claude & GPT in Long-Horizon Tasks — 机器之心 · 2026-08-14
• Dots3-note open-sourced: 280B-total/16B-active multimodal model scores perfect 42/42 on IMO 2026, 1/10 the size of rivals — xiaosun86 · 2026-08-14
• RedNote Releases dots3-note Preview: 280B Open-Source Multimodal Model — AdinaYakup · 2026-08-14
• dots3-note Preview Open-Sourced: 280B Parameters Rivals DeepSeek-V3 — ChengleiSi · 2026-08-14
• Open-Sourcing dots3-note: 280B Multimodal Model & TEMPO RL Framework for Long-Horizon Agents — teortaxesTex · 2026-08-14
• RedNote Team Releases dots3-note: 280B MoE Multimodal Model with 512K Context — bookwormengr · 2026-08-14
• [source] Xiaohongshu Open Sources dots3-notePreview, Proposes TEMPO Training Paradigm — 小红书技术REDtech · 2026-08-14
• Xiaohongshu Open-Sources dots3-note: Perfect 42/42 at IMO, Apache 2.0 — eyishazyer · 2026-08-14
• Xiaohongshu Open-Sources 'dots3-note' Model That Scored Perfect at IMO — eyishazyer · 2026-08-14
• dots3-note: 280B MoE Model with TEMPO for Long-Horizon Agents — iamrobotbear · 2026-08-15