Xiaohongshu's Dots Model Lab open-sourced dots3-note Preview today — weights on Hugging Face and GitHub, Apache 2.0, architecture already submitted to Transformers for day-one library support. Xiaohongshu is RedNote, the lifestyle-and-social app — not a company anyone expected to show up in frontier open-weights model releases, though its Dots team has shipped dots.llm and dots.ocr before this is its first entry into the reasoning-and-agent tier. Note is the smallest of a planned three-model dots3 family (note, jazz, aria) — 280B total parameters, 16B active, 512K context, multimodal across text, vision, and speech.
The headline claim, and the precision it needs
Buried in a section about self-evaluation rather than the launch's top line: at IMO 2026 last month, an internal harness built around a branch of dots3-note Preview generated proofs recursively, used tools to evaluate and improve them, and earned an officially certified perfect score of 42/42 — an IMO gold medal. That "certified" matters; it means the result went through actual competition grading, not a self-reported claim. It's also worth reading precisely: this was "a branch of" the preview model, run inside a purpose-built harness for recursive proof generation — not necessarily bit-identical to the checkpoint now sitting on Hugging Face. The gap between "the capability that won gold" and "the artifact you can download" is the same gap we've flagged in other releases this month, and it's worth checking the technical report, due within a week, for exactly what differs.
TEMPO: a real answer to why long-horizon RL keeps breaking
The more durable contribution here might be methodological. Standard actor-critic RL (PPO and similar) struggles on long-horizon agent tasks for a specific reason the team states plainly: a single rollout can run more than ten hours, making training prohibitively slow, while a critic that estimates value through a fixed-compute forward pass can't reason, reflect, or use tools to judge a complex intermediate state the way the actor can.
Their fix, TEMPO (Test-time-scaled Value Estimation with Macro-step Policy Optimization), decomposes a long task into macro-steps. At the end of each one, the same agent switches roles from actor to critic, using test-time-scaled reasoning — not a separate, cheaper model — to estimate the expected remaining return before the task is even finished. The policy updates on that estimate rather than waiting out the full ten-hour rollout. On ARC-AGI-3, TEMPO scores 31.5% higher than the baseline checkpoint and 20.6% higher than GRPO, and reaches a given score in fewer steps. The team's own worked example is a clean illustration: two training trajectories on a constraint-satisfaction puzzle both ran 64 rounds with identical environment scores, indistinguishable by outcome alone — but one had genuinely found the puzzle's hidden rule and one had latched onto a wrong hypothesis, and TEMPO's self-critique step told them apart. That's a real demonstration of evaluation being easier than generation, the same asymmetry this month's harness coverage has been circling from other directions — one day after Z.ai reported a similarly surprising jump in a different self-evaluation-adjacent capability.
Continual learning inside ARC-AGI-3, and outside its training distribution
Separately from the IMO result, the team validated test-time memory formation using ARC-AGI-3 — the same ultra-long-horizon benchmark this month's harness posts keep returning to — where tasks run thousands of interactions over 40–50 hours and success requires learning the environment's rules with no prior knowledge, the way a human would explore an unfamiliar game. dots3-note Preview reportedly forms hypotheses, tests them through interaction, corrects stale memory when a hypothesis fails, and carries working memory forward into later decisions. The notable claim: this generalized to environments the model was never trained on, demonstrated with a public build running continuously against Slay the Spire 2. If that holds up under scrutiny, it's evidence of a genuinely transferable exploration strategy rather than memorized behavior for one benchmark family.
The benchmark charts, read carefully
Across two dense comparison grids — Personal Assistant Agent, Coding & Computer Use, Information Retrieval, and Complex Reasoning in one; multimodal perception in the other — dots3-note Preview is measured against Hy3 (295B/21B), GLM 5.2 (743B/39B), DeepSeek-V4-Flash-0731 (284B/13B), Kimi K3 (2.8T/104B), Claude Opus 4.8, and GPT-5.5. The team's own framing is accurate to what the charts show: dots3-note Preview is competitive with, and on several benchmarks ahead of, models several times its own size — a 280B/16B model trading blows with a 2.8T-total competitor is a genuinely notable efficiency result, in the same vein as this month's other small-model coverage. It does not lead everything — the closed frontier (Opus 4.8, GPT-5.5) and the largest model shown (Kimi K3) still win the hardest reasoning categories — but "several times its size" is not marketing exaggeration here.
Two new benchmarks worth noting on their own
The release includes open-sourced eval environments, not just a model: VibeSearchBench (200 tasks, 20 domains, multi-turn search where intent is revealed gradually by a persona-driven simulator, scored by knowledge-graph triplet matching) and VibeLifeBench (20 tasks, 10 domains, each spanning 20–30 stages with 1,247 atomic checks for cross-stage consistency, tool execution, and final deliverables — the kind of long-running, state-changing task a real personal assistant would actually face). Both are aimed at the gap between closed, verifiable RL environments and ambiguous real life, and both are usable by anyone trying to reproduce or challenge the results — a genuine contribution to the shared eval ecosystem, not just a model drop.
The limitations section earns real credit
The team states plainly: "Reinforcement learning is not yet complete," with named limitations in hallucination mitigation, the text/multimodal capability balance, and overall stability. They're equally direct that the real-life task demos — wedding coordination, e-commerce operations, a travel concierge — run in simulated environments only, and that turning the capability into something reliable "requires not only a foundation model, but also robust harnesses, connectors, data sources, safety and permission mechanisms, and product design. This work is ongoing." That's the same disclosure standard credited elsewhere this month — stating a capability's actual boundary rather than letting a strong headline result imply more than it proves.
What to expect next
- Watch for the technical report within the week, specifically for how the IMO-winning configuration relates to the released checkpoint — harness details, whether it's the same weights, and reproducibility.
- Watch for independent verification of the size-adjusted benchmark claims, particularly against Kimi K3 and the closed frontier, now that weights are actually downloadable.
- Watch whether jazz and aria — the larger dots3 siblings — follow, and on what timeline. A credible three-tier family, not just a single lightweight preview, is the test of whether Xiaohongshu is building a durable model program or a one-off research showcase.
References: dots3-note announcement · dots-studio/dots3-note-prev on Hugging Face · GitHub — studio-dots-ai/dots3-note-prev · announcement thread on X · related coverage: GLM-5.3 · Understanding ARC-AGI · Nemotron 3.5 Lightning · Qwen3.8-Max's open weights land · Frontier Arcade: trends & predictions