ai lab

Dots Studio

Dots Studio matters because it connects the data and product context of a large social platform with a model program that publishes weights, technical reports, and unusually detailed intermediate artifacts. Its most distinctive work spans inspectable language-model training and compact document models, while dots3-note tests whether that research line can become a competitive multimodal agent system.^2,5,8,10,13

Models in the Atlas →

Profile

Overview

RedNote creates a public model lab

Dots Studio is the public model-publishing identity associated with Xiaohongshu, the Shanghai company whose social and commerce platform is also known as RedNote. Xiaohongshu launched the studio from its in-house language-model team in early 2025, and the Humane Intelligence Lab released the first public weights, dots.llm1, that June. In April 2026 the parent company elevated Hi Lab into Dots, a first-level internal AI department. Dots Studio is therefore the public release identity, while Dots is the organization inside Xiaohongshu.^1,2,3,4

Training checkpoints become part of the release

dots.llm1 was a 142-billion-parameter mixture-of-experts language model that activated 14 billion parameters for each token. The team trained it on 11.2 trillion tokens without synthetic data and released intermediate checkpoints at one-trillion-token intervals. That unusual checkpoint series made training dynamics inspectable rather than publishing only the final weights, although the model's benchmark claims initially came mainly from its own technical report.^5,6

Language expands into documents, vision, and speech

The lab then broadened from language into perception. dots.vlm1 combined a 1.2-billion-parameter vision encoder with a DeepSeek V3 language backbone. dots.ocr used one vision-language model to perform layout detection, text recognition, table extraction, and reading-order reconstruction. Its successor, dots.mocr, expanded the representation to graphics, charts, diagrams, and SVG output, treating visual elements as structured content instead of discarded crops. The lab also released dots.tts, a continuous autoregressive speech model with training and inference code.^7,8,10,12

Dots Studio opens a multimodal agent model

After Xiaohongshu elevated Hi Lab into the Dots department in April 2026, the Dots Studio publishing identity released dots3-note Preview that August. The publisher reports a 280-billion-parameter multimodal mixture-of-experts model with 16 billion active parameters and a 512,000-token context window. It accepts text, images, video, and audio and is designed for long-horizon tool and agent tasks. Apache-licensed weights, FP8 checkpoints, and documented vLLM and SGLang serving paths make the release inspectable, but its benchmark record remains publisher-reported: ModelCap has admitted no measured evidence, while BenchLM maps source-linked rows from the model card. The model is therefore a serious new artifact whose competitive position remains provisional.^13,14,15,16

Notable contributions

01One-trillion-token training checkpointsFor dots.llm1, the team released intermediate checkpoints after each trillion training tokens, alongside the base and instruction models. The contribution is an unusually granular public record of large mixture-of-experts training rather than a claim to have invented checkpointing.^5,6
02Unified multilingual document parsingdots.ocr placed layout detection, content recognition, table and formula extraction, and reading-order reconstruction inside one compact vision-language model, then introduced the XDocParse benchmark across 126 languages.^8,9
03Documents that preserve graphics as codedots.mocr treats charts, diagrams, tables, icons, and text as connected document content and converts visual structures into reusable textual or SVG representations instead of returning only cropped images.^10,11
04An open multimodal model for long-horizon tasksdots3-note combines text, image, video, and audio inputs with long context, tool use, and open weights. Its contribution is the released combination and serving stack; independent evidence is not yet sufficient to validate the publisher's full performance claims.^13,14,15

Sources · 16+

Model lab #28

As of Aug 29, 2026

Dots Studio has real release momentum, while dots3-note still lacks independently measured benchmark evidence.

Dots Studio has moved quickly from language-model training research into OCR, multimodal document parsing, speech, and a new omni-input agent model. dots3-note publishes Apache-licensed weights, FP8 checkpoints, and documented vLLM and SGLang serving paths. The caution is measurement: ModelCap has admitted no measured evidence and labels its index estimate as modeled, while BenchLM maps source-linked publisher results rather than independently reproduced evaluations. The editorial read is positive on breadth and release momentum, but provisional on frontier capability.

Why this editorial standing+

Dots Studio rises from its seed score of 49 to 60 because the earlier record omitted a sustained research sequence across language, vision, document, speech, and multimodal agent models. The new dots3-note release adds current open-weight momentum, but its strongest results remain publisher-reported and ModelCap has admitted no measured evaluation evidence for the model. Within general model labs, Dots Studio therefore ranks above thin or inactive programs but below labs with mature third-party frontier evaluation and broader developer ecosystems.^{2,5,8,10,13,15}

Model evidence: 20 / 35
Research influence: 12 / 25
Release momentum: 15 / 20
Ecosystem: 7 / 10
Trajectory: 6 / 10

Notable people

Revision history1 updates +

Initial researched Dots Studio dossier
Aug 29, 2026
Reconstructed the RedNote lab's language, vision, document, speech, and agent-model history, then raised its score for current momentum while keeping the dots3 capability judgment provisional.