Dots Studio

dots studio was launched by Xiaohongshu in early 2025, growing out of the company's internal large model technology team.^1, 2

Its stated mission is to "Create frontier intelligence for daily life".¹

The dots.llm1 series, released on 6 June 2025 under the MIT licence, is a mixture-of-experts model that activates 14 billion of 142 billion parameters.^2, 4, 5

dots.ocr, a multilingual document parsing model based on a 1.7B language model, was released on 30 July 2025, and dots.ocr-1.5 was rebranded as dots.mocr on 19 March 2026.⁶

dots.vlm1, which the studio describes as its first multimodal large model, combines a 1.2B-parameter vision encoder with the DeepSeek V3 language model.^1, 7

In August 2026 it open-sourced dots3-note preview, a mixture-of-experts model with 280B total and 16B active parameters, a 512K context window and Apache 2.0 weights.^2, 4, 8

The dots3 series is planned to include note, jazz and aria.^2, 8

The studio organises its research around pre-training, multimodal, post-training, AI infrastructure and AI systems.¹

Its dots API open platform, whose service terms took effect on 11 August 2026, offers OpenAI Chat Completions and Anthropic Messages compatible endpoints and, as of October 2026, is offered as a free trial.^9, 10

The API is operated by a Shanghai company, and its privacy policy says personal information collected in China is stored in China.³