On August 14, dots studio — the AI lab under Xiaohongshu (RedNote) — published the first open-weight model of its dots3 family on Hugging Face: dots3-note preview, a 280B-total/16B-activated MoE with a 512K-token context window, released under Apache 2.0. A content platform stepping into open-weight foundation models is itself worth a closer look.

Core Specs: A "Big Total, Lean Activation" MoE

According to the official model card, dots3-note preview is configured as follows:

  • Architecture: Multimodal MoE with 1 dense layer + 45 MoE layers, hidden size 5120;
  • Experts: 256 routed experts + 1 shared expert, top-8 routing; 280B total parameters, 16B activated per token;
  • Attention: 13 DSA layers (Top-2048) + 33 sliding-window attention (SWA) layers, roughly a 1:3 ratio;
  • Context: 512K tokens, 152K vocabulary;
  • Multimodal inputs: A MoE ViT vision encoder (7B total, 1.2B activated) and an 800M dense audio encoder; accepts text, image, video, and audio input, produces text output;
  • MTP: 1 shared layer (1.13B) for speculative decoding;
  • Precision: BF16 / FP8.

The model is positioned as "the most lightweight member of the family" — the dots3 family is designed to include multiple models with different trade-offs among capability, latency, and inference cost, and note preview is simply the first card played.

Self-Reported Evaluations: Coding and Multimodal Both Covered

Evaluation numbers submitted to the Hugging Face model card include: SWE-bench Verified 78.4, SWE-bench Multilingual 75.7, SWE-bench Pro 61, MMMU Pro (standard 10-option) 79.1, WildClawBench 61.7, Claw-Eval 73.4, HLE 52.6, and Apex Agents 30.8.

One caveat worth stressing: these are all self-reported numbers, and third-party reproduction will take time. But the distribution tells a clear story — the narrative centers on coding agents and multimodal understanding, not on pure general-reasoning leaderboard climbing.

Deployment: Engineering Coordination Was Done Up Front

The FP8 checkpoint runs on a single 8-GPU node. vLLM main already has native support, while the SGLang (#33829) and Transformers (#47844) PRs are still under review; the team simultaneously shipped an SGLang Docker image and a vLLM recipe. MTP/NEXTN speculative decoding is optional and, per the model card, can reduce TPOT by more than 50%. The weights are also on ModelScope, with a free entry point on OpenRouter.

Having the inference frameworks moving on day one means this is not a "dump the weights and hope" release — the engineering coordination was done in advance.

A Few Judgments

First, the 16B activation is the real selling point. 280B total parameters sounds intimidating, but MoE inference cost is determined by activated parameters, and 16B puts this model in mid-tier pricing territory. Combined with the 512K context and multimodal inputs — and with tool use, multi-step agent workflows, and interactive tasks explicitly listed in the model card — this is a model designed as an interface for the agent era.

Second, both words in "note preview" deserve scrutiny. Family's lightest member + preview build means dots studio is probing with a small cup: watch community feedback and inference-stack compatibility first, then decide how to release the bigger variants.

Third, a content platform building foundation models holds a card others don't: the corpus. Xiaohongshu sits on massive, authentic multimodal data — images, text, and video — and the fact that the vision encoder is itself a MoE structure suggests it was built for heterogeneous inputs like these. Apache 2.0 licensing plus a free OpenRouter endpoint pushes the developer's cost of trying it to nearly zero.

So What

For the sequence of Chinese open-source models, this is just one more line item. But for the question of "who gets to build open multimodal foundations," Xiaohongshu's answer is: content platforms with authentic multimodal corpora can take a seat at the table too. For developers, running a few of your own agent workloads through the free OpenRouter preview is now the cheapest possible validation — until third-party reproduction numbers arrive, your own workload is the best benchmark.

References: Hugging Face model card · GitHub repository