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 Xiaohongshu's dots model lab, and belongs to the same series as the IMO 2026 perfect-score model. The model adopts a MoE architecture with 280B total parameters/16B activated parameters, and supports a 512K ultra-long context along with text, visual, and speech multimodal understanding. The model is optimized for complex reasoning, Agents, and multimodal perception, specializing in long-horizon open-ended tasks such as travel planning and wedding preparation, and is comparable on multiple reasoning and agent benchmarks to large models with several times the parameters.

Main features of dots3-note preview

• Long-horizon open-domain Agent: Specializes in complex tasks with no single answer and long time spans, such as travel planning and wedding preparation, advancing independently and delivering results.
• Complex reasoning and mathematical proof: Inherits the IMO 2026 perfect-score capability, with top-tier mathematical reasoning and formal proof proficiency.
• Multimodal understanding: 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 engineering: Can independently complete the full development process from requirements analysis, technology selection, code implementation, and compilation and building to verification and running.
• Autonomous learning and error correction: In environments with unknown rules, through observation-based hypotheses and verification feedback, when errors occur it initiates Self-Critiquing correction and writes it into dynamic memory.
• Interactive task execution: Without specialized training, it can take over tasks such as games and cracking ARC-AGI that require real-time feedback and long-horizon strategy adjustment.

Technical principles of dots3-note preview

• MoE mixture-of-experts architecture: The model has 280B total parameters and only 16B activated parameters, adopts a design of 256 routed experts plus 1 shared expert, activates experts per layer through Top-8 routing selection, 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 cost.
• Multi-Token prediction acceleration: Built in is a 1.13B-parameter shared multi-Token prediction layer, which can predict up to 3 subsequent tokens in parallel while generating the current token, enabling speculative decoding without an independent draft model, effectively reducing inference latency and improving throughput.
• Multimodal encoder fusion: Visual information is processed by a MoE ViT encoder with 7B total parameters/1.2B activated, audio is encoded by an 800M-parameter dense network, and features from each modality are projected and aligned with the language model, achieving unified understanding and cross-modal reasoning across text, images, video, and audio.
• Long-horizon task cognition mechanism: Relying on a 512K ultra-long context window, the model in an open environment proposes and verifies hypotheses through observation, initiates Self-Critiquing self-criticism when errors occur, and writes the corrected rules into the external memory.md dynamic memory file, achieving knowledge accumulation and strategy adjustment across steps.

Follow WeChat and reply "open source" to joinAI Open Source Project Exchange Group

How to use dots3-note preview

• Visit the official platform: Go 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 the model weights: Find it on Hugging Face dots-studio/dots3-note-prev repository, download the model weight files under the Apache-2.0 license to the local server.
• Install the inference framework: Select and install a compatible framework such as vLLM, SGLang, or Transformers according to the hardware environment, and load the model to prepare for inference.
• Configure the parallelism strategy: For the 280B total parameter scale, set up tensor parallelism (TP) and expert parallelism (EP) to distribute computation across multiple GPUs to meet VRAM and performance requirements.
• Launch task invocation: Input multimodal data such as text, images, and audio through an API or local interface to trigger long-horizon Agent tasks, complex reasoning, or end-to-end software engineering.

Core advantages of dots3-note preview

• Extreme parameter efficiency: With 280B total parameters but only 16B activated, it achieves reasoning and Agent capabilities comparable to or even surpassing models with several times more parameters, at a computational cost far lower than large models of similar performance.
• Long-horizon open-domain Agent: Unlike standardized scenarios such as mathematics/code, it specializes in real complex tasks with no single answer and long time spans, such as travel planning and wedding preparation.
• Top-tier reasoning genes: Inherits mathematical reasoning and formal proof capabilities with an official perfect score (42/42) in IMO 2026.
• Native multimodal understanding: Uniformly processes text, images, video, and audio, supporting visual spatial interpretation, document and chart parsing, and cross-modal reasoning.
• Autonomous learning and error correction: Built-in Self-Critiquing mechanism automatically corrects errors and writes rules into memory.md dynamic memory, enabling cross-step knowledge accumulation.
• End-to-end engineering capability: Can independently complete the full software engineering process from requirements analysis, technology selection, and code implementation to compilation and verification.

Project address of 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 competitors

Application scenarios of dots3-note preview

• Intelligent renovation consultant: The user uploads a floor plan and appliance dimension drawings and describes requirements by voice; the model automatically calculates spatial suitability, retrieves style cases, generates a customized renovation plan webpage, and reminds for on-site re-measurement.
• Independent Game Boosting Agent: Takes over strategy games such as Slay the Spire II, learns enemy mechanics based on the real-time battle situation, dynamically manages the deck, gold, and potions, and makes globally optimal decisions in long-horizon matches.
• Full-Process Wedding Planner: From budget allocation, venue screening, and guest list management to creating the process timeline, the model continuously follows up over several weeks and dynamically replans based on sudden changes (such as weather or personnel adjustments).
• visionOS Native App Development: Independently completes the requirements analysis, SwiftUI+RealityKit code writing, Xcode project building, and simulator verification for an Apple Vision Pro app based on reference design drawings, delivering a spatial app from 0 to 1.
• Cross-Modal Deep Research Assistant: For academic or business topics, autonomously performs multiple rounds of search, reads PDF charts, watches video materials, integrates multi-source information to generate structured research reports, and initiates Self-Critiquing to revise conclusions when contradictions are found.