dots3-note preview - Xiaohongshu's open-source multimodal MoE model

What is Dots3-Note Preview?

Dots3-note preview is the first version of the Dots3 series, open-sourced by the Dots Model Lab at Xiaohongshu, and is part of the same series as the model that achieved a perfect score at IMO 2026. The model uses a MoE architecture with 280B total parameters and 16B activation parameters, supporting 512K ultra-long context and multimodal understanding of text, vision, and speech. Optimized for complex reasoning, agents, and multimodal perception, the model specializes in long-term open tasks such as travel planning and wedding preparation, and rivals larger models with several times the parameters on multiple reasoning and agent benchmarks.

Main functions of dots3-note preview

• Long-range open domain agentSpecializes in complex tasks with no single answer and long timeframes, such as travel planning and wedding preparation, and independently drives and delivers results.
• Complex Reasoning and Mathematical ProofIt inherits the ability to achieve a perfect score in IMO 2026 and possesses top-level mathematical reasoning and formal proof skills.
• Multimodal understandingIt supports text, image, video, and audio input, and can perform visual spatial interpretation, document and chart parsing, and video question answering.
• End-to-end software engineeringAble to independently complete the entire development process from requirements analysis, technology selection, code implementation, compilation and building to verification and operation.
• Self-directed learning and error correctionIn an environment with unknown rules, it observes hypotheses, verifies feedback, and initiates Self-Critiquing to correct errors and write them into dynamic memory when errors occur.
• Interactive task executionIt can take over games and crack ARC-AGI and other tasks that require real-time feedback and long-term strategy adjustments without special training.

The technical principles of dots3-note preview

• MoE Hybrid Expert ArchitectureThe model has a total of 280B parameters, with only 16B activation parameters. It employs a design with 256 routing experts and 1 shared expert. Each layer selects activation experts through Top-8 routing and is equipped with a 5,120-dimensional hidden layer and a 152K vocabulary, achieving the expressive power of a large-capacity model while controlling inference costs.
• Multi-Token Prediction AccelerationThe built-in shared multi-token prediction layer with 1.13B parameters can predict up to 3 subsequent tokens in parallel while generating the current token. It can achieve speculative decoding without the need for a separate draft model, effectively reducing inference latency and improving throughput.
• Multimodal encoder fusionVisual information is processed by a MoE ViT encoder with 7B total parameters and 1.2B activation, and audio is encoded by a dense network with 800M parameters. The features of each modality are projected and aligned with the language model to achieve unified understanding and cross-modal reasoning of text, images, video, and audio.
• Long-term task cognitive mechanismLeveraging a 512K ultra-long context window, the model proposes and verifies hypotheses through observation in an open environment. When errors occur, it initiates self-criticism, writing the corrected rules into the external memory.md dynamic memory file, thus achieving cross-step knowledge accumulation and strategy adjustment.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use dots3-note preview

• Visit the official platformGo to https://dots.ai/platform/ to register an account and obtain an API key, and directly call the model through the cloud interface without local deployment.
• Download model weightsFound on Hugging Face dots-studio/dots3-note-prev Download the model weight files under the Apache-2.0 license to the local server from the repository.
• Install the reasoning frameworkSelect and install a compatible framework such as vLLM, SGLang, or Transformers based on the hardware environment, and load the model to prepare for inference.
• Configure parallel strategyFor a total parameter scale of 280B, tensor parallelism (TP) and expert parallelism (EP) are set up to distribute the computation across multiple GPUs to meet the requirements of video memory and performance.
• Start task call: Input multimodal data such as text, images, and audio through API or local interface to trigger long-term agent tasks, complex inference, or end-to-end software engineering.

The core advantages of dots3-note preview

• Extreme parameter efficiencyWith a total of 280B parameters, only 16B are activated, achieving inference and agent capabilities comparable to or even surpassing those of models with several times the number of parameters with far lower computational power consumption than large models of the same performance.
• Long-range open domain agentUnlike standardized scenarios such as mathematics/code, it specializes in real-world, complex tasks with no single answer and long time spans, such as travel planning and wedding preparation.
• Top-notch reasoning genesIt inherits the mathematical reasoning and formal proof capabilities that achieved a perfect score (42/42) in the IMO 2026 official certification.
• Native multimodal understandingIt unifies the processing of text, images, videos, and audio, and supports visual spatial interpretation, document and chart parsing, and cross-modal reasoning.
• Self-directed learning and error correctionIt has a built-in Self-Critiquing mechanism that automatically corrects errors and writes the rules to memory.md for dynamic memory, enabling knowledge accumulation across steps.
• End-to-end engineering capabilitiesAble to independently complete the entire software engineering process from requirements analysis, technology selection, code implementation to compilation and verification.

The project address for dots3-note preview

• Project official website：https://studio.dots.ai/dots/dots3-en.html
• GitHub repository：https://github.com/studio-dots-ai/dots3-note-prev
• HuggingFace model library：https://huggingface.co/dots-studio/dots3-note-prev

Comparison of dots3-note preview with similar competing products

Application scenarios of dots3-note preview

• Smart Home Renovation ConsultantUsers upload floor plans and appliance dimensions, and describe their needs via voice. The model automatically calculates space adaptability, searches for style examples, generates a customized renovation plan webpage, and reminds users to take measurements on-site.
• Indie game boosting agentTake over strategy games such as Slay the Spire II, learn enemy mechanics based on real-time battles, dynamically manage decks, gold, and potions, and make optimal global decisions in long matches.
• Wedding plannerFrom budget allocation, venue selection, guest list management to process schedule development, the model continuously tracks progress over several weeks and dynamically replans based on unforeseen changes (such as weather or personnel adjustments).
• visionOS native application developmentBased on the design drawings, independently complete the requirements analysis, SwiftUI+RealityKit code writing, Xcode project building and simulator verification for the Apple Vision Pro application, and deliver the spatial application from scratch.
• Cross-modal deep research assistantFor academic or commercial topics, it independently performs multiple rounds of searches, reads PDF charts, and watches video materials to integrate multi-source information to generate structured research reports, and initiates self-critiquing to revise conclusions when contradictions are found.