dots3-note preview – In-Depth Review of Xiaohongshu's Open-Source Multimodal MoE Model

1. What is dots3-note preview

dots3-note preview is an open-source multimodal MoE model developed by Xiaohongshu's dots model lab. As the first version of the dots3 series, it shares the same technical lineage as the IMO 2026 full-score model. The model employs a MoE architecture with a total of 280B parameters and 16B activated parameters, supporting a 512K ultra-long context and multimodal understanding of text, vision, and audio. It is specifically optimized for complex reasoning, agent tasks, and multimodal perception, excelling in long-horizon open-domain tasks such as travel planning and wedding preparation. In multiple reasoning and agent benchmarks, it can rival large models with several times more parameters.

Image source: Official article
Image source: official article

Technical positioning and domain: Belongs to the intersection of multimodal large models and agent systems, positioned as a long-horizon open-domain Agent model. It differs from traditional Q&A or code generation models by emphasizing autonomous planning, execution, and error correction in real-world scenarios with no standard answers, multiple steps, and long time spans.

Development background: Developed by Xiaohongshu's dots model lab, based on the technical achievements of the same series that achieved a perfect score (42/42) in IMO 2026. The goal is to explore the potential of high-parameter-efficiency MoE architectures in complex real-world tasks and to advance the open-source community's technical progress in the Agent domain.

Core value: Addresses the weaknesses of existing large models in handling real-world tasks that involve multiple steps, long time spans, and no single correct answer—such as poor planning capabilities, limited memory, and inability to continuously correct errors. It achieves cross-step knowledge accumulation and adaptive strategy adjustment through a Self-Critiquing mechanism and external dynamic memory files.

Technical features: Utilizes a MoE architecture with 256 routing experts plus 1 shared expert, achieving the expressive power of a 280B parameter model by activating only 16B parameters; includes a 1.13B parameter mult-Token prediction layer to accelerate speculative decoding; visual processing is handled by a MoE ViT encoder with 7B total parameters and 1.2B activated parameters, while audio is encoded using an 800M dense network, enabling native unified understanding across multiple modalities.

2. Key Features

• Long-horizon Open-Domain Agent: Specializes in complex tasks with no single correct answer and long time spans, such as travel planning and wedding preparation. The model can independently advance and deliver complete results, with capabilities in multi-step planning, resource coordination, and dynamic adjustment, distinguishing it from Agent systems that only handle standardized tasks.

• Complex Reasoning and Mathematical Proof: Inheriting the official full-score capability for the IMO 2026 competition, the model demonstrates top-tier mathematical reasoning and formal proof abilities. It can provide rigorous derivation processes for competition-level problems in number theory and combinatorics, making it suitable for research assistance and educational applications.

• Multimodal Understanding: Supports input from four modalities: text, image, video, and audio. It can perform visual spatial reasoning, document chart analysis, and video question answering. The fusion design of the MoE ViT encoder and dense audio encoder ensures high accuracy and efficiency in cross-modal reasoning.

• End-to-End Software Engineering: Capable of independently completing the full development process, from requirement analysis, technology selection, code implementation, compilation and building, to verification and execution. It can deliver functional applications without human intervention, with typical use cases including visionOS native app development and Web application construction.

• Autonomous Learning and Error Correction: In unknown rule environments, the model can observe, formulate hypotheses, and validate them. When errors occur, it activates the Self-Critiquing mechanism to refine and update rules in the external memory.md dynamic memory file, enabling knowledge accumulation and strategy adjustment across steps, significantly enhancing robustness in long-horizon tasks.

• Interactive Task Execution: Can take over tasks requiring real-time feedback and long-term strategy adjustments, such as gaming and solving ARC-AGI puzzles, without specialized training. The model dynamically adjusts its behavior based on the current state, demonstrating decision-making capabilities comparable to those of human players in strategy games.

3. How to Use

1. Environment Requirements and Prerequisites: Due to the model's total parameter count of 280B, it is recommended to deploy using multi-GPU setups. Each GPU should have at least 80GB of memory (such as NVIDIA A100 or H100). The operating system should be Linux, and CUDA 12.x or higher should be installed. If you are only using the model via API, local hardware is not required.

2. Access the Official Platform (Cloud Inference): Go to https://dots.ai/platform/ to register an account and obtain an API key. You can directly call the model through RESTful APIs, which support multimodal input and long-running Agent tasks. This is ideal for quick trials and lightweight integration, without needing to worry about underlying deployment.

3. Download Model Weights (Local Deployment): Download the model weight files from the Hugging Face repository dots-studio/dots3-note-prev under the Apache-2.0 license, which include model parameters, tokenizer, and configuration files. The weight files are approximately 560GB (FP16), so sufficient storage space should be reserved.

4. Install the Inference Framework and Configure Parallelism Strategy: Choose a compatible framework such as vLLM, SGLang, or Transformers based on your hardware environment. When loading the model, set up Tensor Parallelism (TP) and Expert Parallelism (EP) to distribute the 280B parameters across multiple GPUs. A recommended configuration is TP=8 and EP=4, which can effectively reduce memory pressure on individual GPUs.

5. Initiate Task Inference: Input multimodal data such as text, images, and audio via API or local interfaces to trigger long-running Agent tasks, complex reasoning, or end-to-end software engineering. Note that the first model load will take a considerable amount of time (approximately 10–20 minutes), and subsequent inferences can be accelerated through caching. It is recommended to use streaming output to reduce the first Token latency.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: If the task involves multi-step, real-world scenarios with no standard answers (such as travel planning, wedding preparation, or game strategy), and requires the model to have self-correction capabilities and long-term memory, dots3-note previewis the optimal choice. Its Self-Critiquingmechanism and 512K context windowprovide unique advantages. For scenarios primarily focused on code generation and mathematical reasoning, DeepSeek-v4-flash-0731demonstrates high efficiency with the same activated parameter scale, and its mature community ecosystem is a strong asset.

Further Recommendations: If the user prioritizes multimodal understanding (especially for images and videos) and the tasks are more general conversation-oriented, Qwen2.5-VL-72B, as a dense model, is relatively easier to deploy and offers excellent Chinese support, though it lacks native long-range Agent capabilities. dots3-note previewshows significant differentiation in the Agent domain, making it particularly suitable for researchers and developers exploring next-generation intelligent agent applications.

6. Editor's Summary

The dots3-note preview demonstrates clear innovation in its technical architecture: a MoE design with 256 route experts plus one shared expert, combined with only 16B active parameters, achieves the expressive power of a 280B parameter model. This extreme parameter efficiency provides new insights for the low-cost deployment of large-scale models. The introduction of the multi-Token prediction layer enables speculative decoding without the need for an independent draft model, offering practical value for inference acceleration. The combination of the Self-Critiquing mechanism with external dynamic memory files allows the model to continuously learn and correct errors in long-range open tasks, a critical feature that is currently lacking in most large models.

In terms of practical value, the dots3-note preview exhibits autonomous planning and execution capabilities in real-world complex scenarios such as travel planning, wedding preparation, and game strategies, surpassing traditional Agent systems based on fixed processes. Its end-to-end software engineering capabilities elevate the model from a mere conversational tool to a productive tool capable of delivering tangible results. However, its deployment threshold is relatively high, requiring a multi-GPU cluster to handle the 280B total parameters, which limits its accessibility to individual developers. Additionally, the completeness of its Chinese ecosystem still needs improvement.

This model is suitable for two groups of people: first, researchers exploring agent technology, long-range reasoning, and self-correction mechanisms, as dots3-note preview provides a complete open-source implementation and weights; second, developers who need to handle complex real-world tasks, who can integrate it into applications such as travel planning, home renovation consulting, and game development through API or local deployment. In the future, as the community enhances its support for Chinese, optimizes the inference framework, and releases smaller parameter versions, the dots3-note series has the potential to become a key infrastructure in the open-source Agent domain.

7. Application Scenarios

• Smart Home Renovation Consultant: Users upload floor plans, appliance size diagrams, and describe their needs via voice. The model automatically calculates spatial compatibility, retrieves style case studies, generates customized renovation plan web pages, and reminds users of critical tasks such as on-site measurements, providing a one-stop service from requirement to solution.

• Independent Game Development Agent: Takes over strategy games like Slay the Spire II, learns enemy mechanics based on real-time gameplay, dynamically manages decks, gold, and potions, and makes globally optimal decisions during long sessions. It can complete high-difficulty levels without human intervention.

• Wedding Planning Specialist: From budget allocation, venue selection, guest list management, to creating a timeline for the event process, the model continuously follows up over several weeks. It dynamically replans based on unexpected changes such as weather fluctuations or personnel adjustments, outputting actionable planning documents and reminder lists.

• visionOS Native Application Development: Based on design sketches, the model independently completes requirement analysis, SwiftUI + RealityKit code writing, Xcode project building, and simulator verification for Apple Vision Pro applications, delivering a complete workflow from concept to functional spatial application.

• Cross-modal Deep Research Assistant: For academic or commercial topics, the model autonomously performs multi-round searches, reads PDF charts, and watches video materials, integrating multi-source information to generate structured research reports. When contradictions are detected, it initiates a Self-Critiquing process to revise conclusions, ensuring the accuracy of the final output.

8. FAQ

Q: What are the hardware deployment requirements for dots3-note preview?
A: Local deployment requires a multi-GPU cluster, with 8×NVIDIA A100 80GB or H100 recommended, along with tensor parallelism and expert parallelism configurations. If using only via API calls, no local hardware is required—simply visit https://dots.ai/platform/ directly.

Q: Does the model support Chinese input?
A: The model was primarily trained and evaluated for English scenarios, but the tokenizer includes a Chinese vocabulary, so it theoretically supports Chinese input. However, there is currently limited publicly available performance data for Chinese tasks, so it is recommended to conduct small-scale testing in Chinese scenarios first.

Q: How effective is the 512K context window in practical use?
A: The 512K context window is technically advanced, but its practical effectiveness depends on the task type. For Agent tasks requiring ultra-long memory (e.g., wedding planning spanning multiple weeks), the model can maintain good consistency. However, for random information retrieval, there may be issues with attention dispersion. It is recommended to use external memory files in conjunction.

Q: How can I obtain an API key?
A: Register for an account at https://dots.ai/platform/, then create an API key in the console. The API supports multimodal input and streaming output, and billing is based on token count or number of requests. For specific pricing, please refer to the platform announcements.

Q: Compared to DeepSeek-v4-flash-0731, which is more suitable for code generation?
A: DeepSeek-v4-flash-0731 has been specifically optimized for code and mathematical reasoning, and its community ecosystem is more mature, making it more suitable for pure code generation tasks. dots3-note preview excels in long-range open-domain Agents and multimodal understanding, with strong code generation capabilities, but this is not its primary focus.

Q: Can the model weights be used for commercial purposes?
A: Yes. dots3-note preview is licensed under the Apache-2.0 license, with full open-sourcing of weights and code. It allows for commercial use, modification, and redistribution, but the copyright notice must be retained.

9. Project Links

• Product Website: https://studio.dots.ai/dots/dots3-en.html
• Official Experience Platform (API): https://dots.ai/platform/
• GitHub Repository: https://github.com/studio-dots-ai/dots3-note-prev
• Hugging Face Model Library: https://huggingface.co/dots-studio/dots3-note-prev