The release and what it means
On August 14, 2026, Dots Studio, the model lab of Chinese social platform Rednote (Xiaohongshu), released the model weights for dots3 note preview. The release is the first open-weight model in the dots3 series and belongs to the same series as the model that previously achieved a perfect score of 42 points at the 2026 International Mathematical Olympiad (IMO). The team framed the move as an effort to share its latest technical direction with the wider industry and to encourage work on long-horizon real-world tasks.
Model specifications and capabilities
Dots3 note preview carries 280 billion total parameters with 16 billion active parameters and supports a context window of up to 512K tokens. The model handles multimodal understanding across text, vision, and audio, and has been optimized for complex reasoning and long-horizon agent tasks. On mainstream benchmarks, Dots Studio reported that the model ranks among the leading Chinese models of comparable size in reasoning, agent capabilities, and multimodal perception, with particularly strong visual capabilities for its size, and said it can match or outperform much larger models on several reasoning and agent evaluations.
Why long-horizon real-world tasks
Dots Studio argued that long-horizon real-world tasks will become an important next frontier for large-model capabilities, even though the industry has paid relatively little attention to them. The team contrasted such tasks, like planning a trip, renovating a home, or organizing a wedding that unfold over days or months, with relatively closed-ended tasks such as mathematics and coding where answers and feedback are clearer. To handle real-world tasks, the studio said models must understand non-text information such as floor plans, quotations, flight screenshots, maps, and voice memos, and must continuously evaluate and adjust their plans during execution rather than judging success only at the end.
Technical approach
The technical approach behind dots3 note preview centers on three areas, according to Dots Studio: stronger multimodal understanding to process the diverse information types encountered in everyday scenarios; the introduction of self-critiquing so that long-horizon tasks do not depend solely on final outcomes; and capabilities designed to handle multi-day, multi-step planning that reflects the studio's mission to "Create frontier intelligence for daily life." Two benchmarks introduced earlier by the team, built around complex everyday-life scenarios, had shown leading worldwide models performing poorly and none reaching the passing threshold, reinforcing the studio's case for focusing on this direction.
What remains uncertain
The release material emphasizes benchmark performance and the design philosophy but does not detail training data composition, licensing terms for the open weights, or deployment plans. Independent third-party benchmark reproduction and downstream evaluations were not available in the supplied sources, and any ranking claims between dots3 note preview and competing Chinese models of similar size rest on the studio's own reported results.
Share this article
