Dots.3 Note Preview: Xiaohongshu's New MoE Model, Tested
Dots.3 Note Preview is a 280B MoE multimodal model from Xiaohongshu's AI lab. Here's what its specs, benchmarks, and hands-on tests show.
What is Dots.3 Note Preview?
Dots.3 Note Preview is a new open weight multimodal model from dots.llm, the AI research arm of Xiaohongshu (also known as Red Note), the Chinese social platform often described as a cross between Instagram and Pinterest. It’s a mixture of experts (MoE) model with 280 billion total parameters but only 16 billion active per forward pass, and it’s the first release in a planned Dots.3 model family. It takes text, image, video, and audio as input, outputs text only, and supports context windows up to 512K tokens.
TL;DR
- Dots.3 Note Preview is a 280B-parameter MoE model from Xiaohongshu’s dots.llm lab, activating only 16B parameters per pass while accepting text, image, video, and audio input.
- The model supports a 512K token context window, putting it in range for long documents, extended agent sessions, and large codebases.
- On published benchmark charts covering reasoning, agentic tasks, and multimodal understanding, it reportedly leads its weight class and trails only larger models like Opus 4.5 on the Arc AGI 3 environment-learning benchmark.
- In a hands-on agentic cloud test, the model was given raw AWS credentials and a single goal, and it independently provisioned an EC2 instance, configured a security group, and opened SSH access with no step-by-step guidance.
- A vision-to-code test had the model examine a themed restaurant UI screenshot and rebuild it as an animated web app, correctly reproducing color tinting per food category (chicken, lamb, beef, seafood, vegetarian) rather than just copying layout.
- A role-play stress test showed the model sustaining a coherent, emotionally consistent internal monologue under a forced-choice scenario, sticking to constraints instead of dodging the dilemma.
- A multilingual test asking for a pet nickname translated into dozens of languages with phonetic guides showed strong coverage overall, with the model flagging its own uncertainty on lower-resource languages rather than confidently guessing wrong.
One coffee. One working app.
You bring the idea. Remy manages the project.
What are the core specs?
Dots.3 Note Preview is built as a sparse mixture of experts model: 280 billion parameters total, but a router activates only 16 billion of those for any given token, which keeps inference cost closer to a much smaller dense model while retaining the capacity of a much larger one. That 16B active figure is notable because independent testing found it to be the smallest active-parameter model among its comparison set, yet it still tracked at or near the top on multiple benchmark categories.
The model is multimodal on the input side only. It can process text, images, video, and audio, but it produces text output. Context length tops out at 512K tokens, which is large enough to handle long transcripts, sizable codebases, or multi-turn agent sessions without truncation.
As with earlier dots.llm releases, the model is distributed as open weights, continuing the pattern Xiaohongshu’s lab has followed with its previous model generations.
How does it perform on benchmarks?
Benchmark charts shared alongside the release split into two groups. The first covers reasoning and agentic tasks: personal assistant workflows, coding, long-horizon agent execution, deep search, and competitive programming. Dots.3 Note Preview reportedly leads its weight class across this entire set. The standout result is Arc AGI 3, a benchmark built around learning from an environment rather than pattern-matching on static data. On that test, the model trails only Opus 4.5 while outperforming the rest of the field.
The second chart covers multimodal tasks: vision, documents, video, and audio. Here too, the model reportedly sits at or near the top against larger competitors, including Gemini 3 and Qwen 3.5 Omni Plus, despite activating far fewer parameters than either.
The headline takeaway from these charts is efficiency: a 16B active-parameter model competing with, and in some cases beating, dense and MoE models with substantially larger active footprints.
Can it actually run agentic tasks unsupervised?
One practical test gave the model direct AWS credentials and a single instruction: provision an EC2 instance on a personal AWS account, with no further hand-holding. Using an agent framework (Hermes was used in the test) connected to an AWS CLI, the model worked through the task on its own. It identified the CLI environment, accessed the terminal, used AWS Systems Manager (SSM), and ran the necessary CLI commands.
The result: a running EC2 instance, named “dots3-note-demo,” with a properly configured security group opening port 22 for SSH, plus a public IP address. Refreshing the AWS console confirmed the instance was live and correctly configured. The model was then asked to tear it down, and it terminated the instance cleanly, which the console also confirmed.
This kind of test matters because it goes beyond a chat response. It requires the model to chain together CLI discovery, tool calls, state-checking, and error-free execution against a real cloud environment, with no scripted steps provided.
How does it handle image-to-code and vision reasoning?
A second test fed the model a screenshot of a themed restaurant UI, a rotisserie-style site with tabs for different meats, and asked it to rebuild the interface from scratch as an animated web app, following the original layout and color theme while swapping in categories like chicken, lamb, beef, seafood, and vegetarian.
The output preserved the background tint and overall visual theme of the source image. More notably, the model adjusted color coding per category (green tones for vegetarian, sea green for seafood, and so on) and updated animated flame and smoke effects and titles as the user switched tabs. That level of consistency, correctly inferring and applying a color logic that wasn’t explicitly spelled out in the prompt, points to genuine visual grounding rather than surface-level layout copying.
Is it good at nuanced reasoning and language tasks?
Two further tests probed less technical, more qualitative capabilities.
The first was a forced-choice role-play: the model had to inhabit a character facing an emotionally loaded, no-escape decision between two people, with career and personal consequences either way. The model sustained an in-character internal monologue, honored the constraint that no third option or refusal was allowed, made a definite choice, and produced separate in-character responses to both parties reflecting that choice. It also self-rated its confidence in the outcome, an unprompted touch that read as more honest than performative.
The second was a multilingual test: translate a playful nickname into a long list of languages, each with a phonetic pronunciation guide. The model handled the large majority of entries capably, including major world languages like Turkish, Persian, Polish, and Tagalog, and it flagged uncertainty on genuinely obscure languages like Amharic rather than fabricating confident-sounding nonsense. It stumbled on a few low-resource languages, in one case repeating an Urdu translation for a related-but-distinct language, and it left one term untranslated where a real equivalent exists. The model itself flagged some of these gaps in a closing note, which is a small but useful signal of calibrated self-awareness.
Frequently Asked Questions
What company makes Dots.3 Note Preview?
It comes from dots.llm, the AI research lab operated by Xiaohongshu (Red Note), the Chinese social media and content platform.
What kind of model architecture does it use?
It’s a mixture of experts (MoE) model with 280 billion total parameters, of which 16 billion are active per inference pass.
What’s the context window size?
Up to 512K tokens, which supports long documents, extended agent workflows, and large codebases without hitting truncation limits quickly.
What input and output modalities does it support?
Input includes text, image, video, and audio. Output is text only.
How did it perform in real-world testing?
In hands-on tests it independently provisioned and tore down an AWS EC2 instance from a single goal, rebuilt a themed UI from a screenshot into working animated code, sustained coherent reasoning in a high-pressure role-play scenario, and translated a phrase into dozens of languages with mostly accurate phonetics, flagging its own uncertainty on obscure languages.
