On August 14, 2026, dots studio, the AI research lab inside Rednote (Xiaohongshu, the Chinese lifestyle platform), released the first open-weight model in its dots3 series: dots3-note-preview. The release came alongside two new evaluation frameworks that exposed a gap enterprise AI builders have been circling but not measuring directly: frontier AI models that ace closed benchmarks and mathematical olympiads cannot reliably complete real-world long-horizon tasks.
Every frontier model tested scored below the passing threshold on VibeLifeBench. Claude Opus 5 led at 32.5%. DeepSeek-V4-Pro scored 21.1%. No model demonstrated reliable competence across all ten real-world life domains in the evaluation.
For teams building agentic workflows in production, this is a benchmark result that deserves serious attention.
What Shipped on August 14
dots3-note-preview is a Mixture-of-Experts model with 280 billion total parameters and 16 billion active parameters per token. It supports a context window of up to 512K tokens, processes text, images, video, and audio inputs, and produces text output. Model weights are released under the Apache 2.0 license on Hugging Face and ModelScope.
The architecture uses one dense layer followed by 45 MoE layers with 256 routed experts plus one shared expert (top-8 selection). A separate 7-billion-parameter MoE vision encoder and an 800-million-parameter audio encoder handle non-text modalities. The model runs on vLLM main and is compatible with FP8 and BF16 precision on eight NVIDIA H100 GPUs.
Alongside the model, dots studio released as open source:
- VibeSearchBench: 200 tasks across 20 domains evaluating multi-turn search with progressively disclosed user intent. A persona-driven simulator gradually reveals constraints across conversation turns and compares the agent’s predicted knowledge graph against a ground-truth graph using Triplet F1 as the primary metric.
- VibeLifeBench: 200 multi-week tasks across 10 life domains where the world updates silently on its own clock and the agent must proactively check the environment to discover changes. Jointly developed with Evolvent AI; the 20-task open subset is on GitHub and the full benchmark paper is on arXiv.
The Benchmark That Changes How Enterprises Should Think About Agents
VibeLifeBench is the more consequential of the two new frameworks for enterprise AI teams. It measures three properties that existing evaluations skip entirely.
Proactivity on a live clock. The simulated environment does not wait to be queried. Silent changes, such as flight cancellations, price updates, and booking closures, happen between agent turns. An agent that only responds to prompts will fail because it never learns about changes it was never explicitly notified of.
A world that changes independently of the agent. The simulator maintains a stateful world with 22 mock service backends exposing 288 tool interfaces, driven by 7,453 scripted events. Many events arrive with no notification. Only an agent that actively re-inspects the world catches them.
Long-horizon coherence across a full multi-week lifecycle. The median task in VibeLifeBench spans 29 simulated days. The longest tasks span several months. The benchmark tests whether an agent can keep its plan self-consistent and honor initial constraints throughout a full multi-week trajectory, not just a single session.
The result is unambiguous across all seven models evaluated.
| Model | VibeLifeBench avg@3 | VibeLifeBench max@3 |
|---|---|---|
| Claude Opus 5 | 32.5% | 41.2% |
| Gemini 3.5 Pro | ~30.1% | - |
| Grok 4.x | ~27.5% | - |
| GPT-5.6 Sol | ~27.3% | - |
| Kimi K2.6 | ~25.4% | - |
| dots3-note-preview | ~22.8% | - |
| DeepSeek-V4-Pro | 21.1% | - |
Results from dots studio evaluation report. All models run under the OpenClaw harness with a 256K context window at maximum reasoning settings. Scores are avg@3 on benchmark version v1.0.0. Rows without max@3 reflect dots studio’s summary reporting format.
No model reaches 50%. The benchmark defines a formal passing threshold, and none of the seven evaluated systems clears it. The arXiv paper describes the pattern directly: “contemporary agents, though quite fluent at single-turn tool use, are still far from able to manage life affairs proactively and persistently over weeks in a world that evolves on its own.”
For enterprise AI teams, the implication is direct. If the best available models score 32.5% on carefully structured, clearly defined multi-week tasks in a simulated environment, real-world enterprise deployments on messier, less structured problems will face a steeper version of the same gap. This aligns with what Anthropic disclosed in its August 2026 risk report: internal safety benchmarks saturate while real-world task completion gaps remain wide.
The IMO Connection and What It Reveals
dots3-note-preview is the lightweight member of the dots3 family, which is the same model series that achieved a certified perfect score of 42 out of 42 at the 2026 International Mathematical Olympiad. That result, announced in July, represented the first certified perfect score for any AI system at a major international math competition. The harness enabling it ran iterative self-critique: the model recursively generated proofs, used tools to evaluate them, and refined them until a correct proof was confirmed. The same self-critiquing mechanism is embedded in the dots3-note-preview released now.
A model that can solve every IMO problem correctly but scores 32.5% on multi-week real-world task management is not a contradiction. It is an accurate description of where current AI capability sits. Closed tasks with clear right/wrong answers and rapid feedback respond to scaling and post-training. Open-ended tasks with dynamic environments, implicit constraints, and sparse delayed feedback do not, at least not yet.
This is precisely the frontier the dots studio team is targeting.
TEMPO: Reinforcement Learning for Long-Horizon Tasks
Dots studio developed TEMPO specifically to address the credit-assignment problem in long-horizon RL. Standard reinforcement learning struggles when an agent trajectory runs for dozens of steps across simulated weeks because it is difficult to attribute a final outcome to any individual intermediate decision. Feedback is sparse, and trajectories are long.
TEMPO’s approach:
- Divides long trajectories into macro-steps rather than treating each action independently.
- Trains the model to critique its own intermediate states: identifying mistakes, re-assessing hypotheses, and estimating whether it is on track toward the goal.
- Uses self-critique as a continuous value-estimation signal throughout the trajectory, replacing a single sparse terminal reward.
The same self-critiquing capability that improves training can be deployed at inference time. For enterprise agentic workflows, this means a dots3-based agent can re-evaluate its own intermediate outputs before committing to the next action. That property matters significantly in multi-week deployments where a wrong assumption in day one compounds through day twenty.
What Enterprise AI Builders Should Take From This
The VibeLifeBench results should change how teams evaluate and deploy AI agents for long-running work.
Test against your actual deployment horizon, not just isolated task completion. The Salesforce 2026 Agentic Enterprise Index found that enterprise AI agent deployments have grown rapidly, but most measure task completion on individual, well-defined steps. VibeLifeBench tests multi-week coherence. Even if your team does not run VibeLifeBench directly, its evaluation design provides a practical template: simulate a dynamic world, inject silent changes, and measure whether the agent adapts proactively over time.
Architect for proactivity, not just responsiveness. Most enterprise agent architectures today are prompt-response: a user sends a message, the agent acts, the loop waits. The real-world agentic use cases enterprises care about most, including GTM pipeline monitoring, supply chain exception handling, and long-running contract workflows, require an agent that checks the world between prompts. That is an architectural difference, not a model capability difference.
Evaluate open-weight options for cost-sensitive execution steps. At 16 billion active parameters with Apache 2.0 licensing, dots3-note-preview runs significantly more cheaply than frontier-sized models for intermediate reasoning steps in long agentic workflows. The benchmark results show it scores below frontier models on the overall task, but its cost-per-step economics are relevant for teams designing enterprise inference architectures where the execution layer must scale economically.
Model proactivity and world-inspection as first-class requirements in your agent harness. The VibeLifeBench evaluation runs each agent under the OpenClaw harness, which provides a structured workspace and system prompt that includes access to all 22 mock services. Enterprise teams designing harnesses for long-running agents need to build in the equivalent: structured access to state-check tools, scheduled world-inspection intervals, and alerts on silent environmental changes.
What Comes Next in the dots3 Series
dots studio described dots3-note-preview as the lightest member of a three-tier family: note (released now), jazz (upcoming, mid-tier capability), and aria (upcoming, highest capability). The full dots3-note model with open weights will follow, along with a full technical report.
The competitive context is notable. Xiaohongshu is a social media and lifestyle platform, not a model lab or cloud provider. Its AI team produced a model that earned a perfect IMO score, then released the underlying model as open weights with Apache 2.0 licensing, alongside two new evaluation frameworks that put every major frontier lab on notice about a capability gap their own products share.
For enterprise teams tracking open-weight models production-ready for long-running agentic use cases, dots3-note-preview is worth immediate evaluation.
The dots3-note-preview weights are available on Hugging Face under Apache 2.0. VibeLifeBench’s 20-task open subset runs via GitHub. The benchmark paper is at arXiv 2608.10875.