A 280B Open-Source Model Just Made a Bet: Real AI Agents Need to Run for Hours, Not Seconds
Most model releases lead with a benchmark chart. This one leads with a different pitch: serving real, long-horizon tasks. Xiaohongshu's tech team just open-sourced dots3-note Preview, the lightest model in their dots3 family, and it's built around a specific idea — that the next wave of useful AI isn't about answering one question well, it's about staying on task across dozens of steps.
Here's what shipped, and why the design choice matters.
The Numbers
dots3-note Preview has 280B total parameters, but only 16B are active at any given time — a mixture-of-experts setup that keeps inference cost down while still giving the model a large total capacity to draw on. It supports a 512K token context window and handles text, vision, and speech in one model.
That's a lot of specs. But the headline isn't the parameter count — it's what the team optimized for: complex reasoning and long-running agent tasks, plus native multimodal understanding across all three input types.
Why "Long-Horizon" Is the Interesting Part
Most models are still evaluated on single-turn tasks: answer this question, solve this math problem, write this function. Real agent work doesn't look like that. It looks like: read five documents, call three tools, check the output, correct a mistake, call two more tools, and only then produce a result — sometimes over dozens of steps and long stretches of context.
Models trained and evaluated on short tasks tend to degrade badly the longer a task runs. Context gets lost, earlier instructions get dropped, small errors compound. Building specifically for long-horizon agent tasks means training and evaluating the model on exactly that kind of multi-step, tool-using workflow — not just hoping a strong single-turn model happens to generalize.
That's a meaningfully different design target, and it's part of a broader pattern this month: models are increasingly being built and marketed around staying power, not just raw intelligence.
Multimodal by Default, Not Bolted On
The other notable choice is native support for text, vision, and speech in a single model, rather than stitching together separate encoders after the fact. For agent work specifically, that matters — a lot of real tasks involve reading a screenshot, listening to an instruction, and acting on a document all in the same session. A model that treats all three as first-class inputs from the start tends to hold context across modalities more reliably than one where multimodality was added later.
Open Weights, Real Access
dots3-note Preview is released with open weights, which means developers can actually run it, fine-tune it, and inspect how it behaves — not just call it through an API. That's consistent with a broader trend this year: several of the strongest agent-capable models (DeepSeek, Qwen, GLM, MiniMax) have shipped as open weights rather than staying locked behind a paid endpoint. For teams building agent products, that access matters as much as the benchmark numbers.
What This Means If You Use OpenClaw
The whole premise behind dots3-note Preview — that agents need to hold together across long, multi-step tasks, not just answer one prompt correctly — is exactly the problem OpenClaw is built around.
When you run a tutorial on OpenClaw, your agent isn't graded on one clean response. It's calling tools, checking its own work, picking up context from earlier in the session, and course-correcting when something doesn't go as planned — the same "long-horizon" behavior this new model was purpose-built to handle. The model underneath matters, but so does the scaffolding around it: memory, tool orchestration, and the ability to actually finish a task instead of just starting one.
The Bigger Picture
The trend across this week's releases — long-running agents from xAI, agentic RL training pipelines from Ant Group, and now a long-horizon multimodal model from Xiaohongshu — points the same direction: the AI industry has stopped optimizing purely for "smarter answers" and started optimizing for "can it actually finish the job."
That's the bar OpenClaw agents are built to clear. If you want to see what a long-horizon, tool-using agent looks like in practice, not just in a research paper, you can try it yourself.