Xiaohongshu's dots3-note: a 280B open MoE that only activates 16B

Xiaohongshu's first open-weight model is dots3-note preview, built by its dots studio: a Mixture-of-Experts design with 280B total parameters and 16B activated per token. It accepts text, images, video, and audio as input, writes text, carries a 512K-token context window, and ships under Apache 2.0.

The release is the opening move of a family, not a one-off. The dots3 line is designed around trade-offs among capability, latency, and inference cost, and the card describes dots3-note preview as its most lightweight member. That puts it at the cheap end of a design space the studio left itself room to grow into.

Xiaohongshu's first open weights

Xiaohongshu is not an obvious source for frontier-class weights. It runs a consumer platform, not an AI infrastructure business. The debut lands at a moment when open-weight publication has become the norm for China's AI industry, from Kimi K3's record-size open release on down. The card's job list is broad: instruction following, math and logical reasoning, tool use and multi-step agent workflows, code generation, long-context processing, and understanding of images, documents, charts, audio, and video.

What the card does not publish matters as much as what it promises. The evaluation sections for general reasoning and agents and for multimodal understanding are listed, but no figures sit under them. A full report is marked "coming soon."

A 17:1 bet on sparse parameters

Read the spec sheet and one number stands out. With 16B activated out of 280B total, the ratio of total to active parameters is roughly 17:1, a sharper sparsity than some current open MoE designs. Qwen3-VL, among the most important open-weight video-language families, tops out at a 235B-A22B variant, which is closer to 10:1. At this scale, the cost per token is governed by the 16B figure, not the 280B one. Nvidia's Nemotron 3.5 Lightning pushes the same logic into agent territory, running on 3B active parameters.

The architecture, in brief:

Several serving-relevant details hide in those rows. The mixture runs 256 routed experts plus one shared expert with top-8 routing. The vision encoder is itself an MoE, 7B total with 1.2B active, while the audio encoder is a dense 800M. Attention splits across 13 DSA and 33 SWA layers, about 1:3. A multi-token prediction head, one shared 1.13B layer, feeds an optional speculative decoding path.

The benchmark appendix is empty

This is where anyone weighing the release should slow down. The card advertises a wide capability range, yet as published it offers no evaluation numbers to back it. Agentic claims deserve particular caution: tool use, multi-step workflows, and tasks that require exploration and memory are exactly where the wider field keeps finding that benchmarks flatter the models. A study of tool-using models found the tools can make easy questions harder, and unified agent-loop tests like SceneActBench show even the best VLMs fail at 3D action. Published evaluations have measured agents at 49% on a test human experts pass at 95%. The 512K context window deserves the same skepticism, since advertised context and usable context are rarely the same number.

None of this means the model underdelivers. It means the vendor numbers are not out yet, and comparisons to rival open models should wait until they are.

Running it: FP8, one node, two pending pull requests

The deployment story is more concrete than the benchmark story, with one caveat: support is half-merged. vLLM carries native support on its main branch, so users need a recent nightly build until the next stable release. SGLang and Transformers integrations exist only as pull requests, #33829 and #47844, both still under review. Until they land, the documentation points users at those PR branches.

The recommended path is the FP8 checkpoint on a single 8-GPU node. The vLLM example runs eight H100s with tensor parallelism 8, expert parallelism 8, deep_gemm as the MoE backend, and a context length of 262,144 tokens. That is half the advertised maximum, and the card recommends tuning context to available memory, concurrency, and input modality. BF16 is supported but needs more memory.

The SGLang route ships as a Docker image, lmsysorg/sglang:dev-dots3-note, which pulls the FP8 checkpoint from Hugging Face on first run. Flags cover the fa3 attention backend and an optional speculative decoding setup where the multi-token prediction head runs as NEXTN; the docs say this can cut time-per-output-token by more than 50%. Prefill CUDA graphs are not supported yet.

For lighter testing, the Transformers path adds torchcodec and FFmpeg for audio and video on top of the usual stack. Both serving stacks expose an OpenAI-compatible API, and tool calling turns on with a dots-specific parser flag. A chat template option, enable_thinking, switches between a reasoning trace and a direct response, an option worth having when over half of a model's reasoning is froth.

What a consumer platform wants with open weights

The strategic pattern behind this release is efficiency. MiniMax's M3 already combines native multimodality, a 1M-token context, and agentic coding in an open package; the company has also pushed one architecture across every modality with H3. Qwen spans dense and MoE variants from 30B-A3B to 235B-A22B. On the extreme end, the open 975B-parameter Inkling shipped at 1.9TB before Unsloth compressed it to 270GB at 1-bit with 74.2% accuracy retention. Even small guardrails play the same game: Mistral claims its 3B Shieldstral matches open text-safety models up to seven times its size, as covered in our Shieldstral write-up.

dots3-note preview fits the pattern with an unusual parent. The Apache 2.0 grant keeps the weights usable in commercial products without restriction, and the FP8-on-one-node story suggests the studio is pricing for real deployments, not just leaderboard runs. The same cost pressure shows up across Chinese open weights: GLM-5.2 went viral at $0.07 per million tokens. Because the family is planned around multiple capability and cost points, this is likely the first of several releases.

What is missing is the report. A 17:1 sparsity ratio and a 512K window only become interesting once the numbers behind them are public. Until then, dots3-note preview is an architectural statement from an unexpected quarter. The benchmarks will decide whether it is more.