A 280B-Parameter Lightweight Model Focused on Long-Horizon Agents and Multimodal Reasoning

Key Highlights

Xiaohongshu Technology has open-sourced dots3-note Preview, the lightest model in the dots3 family. Its total parameters are 280B, but only 16B are activated, a typical MoE mixture-of-experts architecture that trades a smaller activation amount for expression power close to a large model while keeping inference cost down. The model focuses on two main lines: long-horizon agent tasks and multimodal reasoning, which together define its personality as a worker rather than a chat toy. In other words, it wants to do both think for a long time and understand what it sees well, which is exactly the combination needed for assistants that operate inside messy real products. The choice to open-source a preview rather than hoard it also signals confidence that the community will find uses the original team did not anticipate, and that feedback will sharpen the next release.

What It Does and How It Unfolds

dots3-note Preview supports a 512K context length, enough to stuff an entire book, a whole codebase, or a long multi-turn conversation into a single inference, so the model rarely loses the thread mid-task. More importantly, it understands text, vision, and voice inputs simultaneously, meaning it can read images, listen to audio, and read text and judge comprehensively, then act on the combined signal. For long-horizon agents, the team made dedicated optimizations so the model is more stable on tasks requiring multi-step planning, repeated tool calls, and maintaining goals across long spans, the kind of work where smaller models usually wander off. Such ability is especially critical for automation workflows, smart customer service, and content moderation, where a single session can involve dozens of dependent steps. By holding context and intent together, the model reduces the fragmentation that breaks most agent loops, making end-to-end automation actually usable rather than demo-only.

Technical Details

The design of 280B total and 16B activated is a typical embodiment of the MoE architecture: each inference wakes only a few expert networks, preserving knowledge capacity while controlling compute cost, because the inactive experts stay dark and cheap. Behind the 512K context lies a mature long-sequence training and positional-encoding scheme that prevents the model from forgetting the beginning of a long input. On the multimodal side, the model aligns image, text, and audio signals uniformly during pre-training, making cross-modal reasoning more natural than late-stage fusion of separate towers. It is worth noting that the Preview label means it is still a preview form that may iterate further based on community feedback, and the open weights let developers try it early instead of waiting for a polished final. That openness also invites scrutiny, which tends to surface bugs faster than a closed beta and improves the eventual stable release through real-world stress.

Comparison With Competitors

In the light yet strong track, dots3-note directly faces competition from open-source models like Qwen, DeepSeek, and Zhipu. The 280B/16B activation ratio balances effect and cost, a similar idea seen in many MoE models, but its specific tuning for agents is the part that stands out. Compared with pure-text models, its multimodal and long-horizon agent orientation is more differentiated; compared with other multimodal models, it keeps the parameter scale in a relatively approachable range that fits on affordable hardware. Put simply, Xiaohongshu wants to use a good enough, cheap, and capable model to prove that a content community can also polish a useful underlying capability, not just consume models made elsewhere. Against giant labs, the edge is proximity to production data: a social platform sees the exact tasks its users need, and can train toward them directly rather than guessing at a generic distribution.

Industry Impact and Use Cases

For domestic developers, there is one more multimodal agent base that can be privately deployed, with room for content understanding, automated operations, and community moderation, all of which a lifestyle platform like Xiaohongshu needs internally every day. For the industry, content platforms building models themselves shows that applications feeding back into models is becoming a trend, as data and needs close to real scenarios are advantages closed-source giants do not necessarily have. A model trained on the messy reality of user-generated content understands that world better than a generalist does. In the long run, lightweight MoE plus long context plus multimodality is likely to become the mainstream recipe for next-generation everyday AI assistants, because it packages the three properties users actually care about into a footprint teams can afford to run. The dot of separation between model maker and app maker is blurring, and that blur is the real story here, because when the platform that understands the use case also trains the model, the feedback loop tightens and the product fits its users far better than a general model ever could, and that is the quiet advantage few competitors can copy.