Open‑Source Dots3‑Note: From IMO Full‑Score Math to Real‑World Long‑Term Tasks

Today XiaoHongShu's Dots Lab announced the open‑source preview of dots3‑note , the first released model in the dots3 series. The model has 280 B total parameters (16 B activation), supports a 512 K token context window, and provides text, vision, and speech multimodal understanding, with optimizations for complex reasoning, agents, and multimodal perception.

Benchmark results show that on several reasoning and agent tasks the model matches or surpasses much larger models, and its visual capabilities stand out within the same size range.

To evaluate real‑world long‑term abilities, the authors tested the model on diverse scenarios:

Slay the Spire II : without task‑specific training, the model learned card and enemy mechanics from combat feedback, managed health, cards, gold, and potions, and reached floor 33 by continuously weighing decisions.

ARC‑AGI 3 : the model discovered synchronization and mirror patterns, identified hazards, movable markers, pressure switches, and gates, and solved the puzzle in 320 steps, using a Self‑Critiquing loop that rewrites corrected rules to a memory.md notebook.

Home‑renovation planning : given a floor plan image, two refrigerator specs, and a partial countertop length, the model calculated remaining wall space, provided a placement recommendation, and suggested on‑site re‑measurement.

VisionOS app development : the model interpreted product requirements and nine UI sketches, chose SwiftUI + RealityKit, generated 12 Swift files (1 876 lines of code), built the Xcode project, and achieved a “BUILD SUCCEEDED” status in the VisionOS simulator, delivering a functional immersive shopping experience.

These demonstrations highlight the model's ability to explore, retain useful information, and self‑correct when judgments deviate, a capability enabled by two key innovations:

Long‑horizon exploration training : thousands of novel, ultra‑long environments without prior knowledge force the agent to continuously learn new rules and retain valuable knowledge beyond the context window, verified on ARC‑AGI 3.

TEMPO (Test‑time‑scaled Value Estimation with Macro‑step Policy Optimization) : trajectories are split into macro‑steps; after each step the model switches from actor to critic, performs inference and tool calls, and estimates future returns to provide intermediate training signals, allowing reinforcement learning to scale to very long tasks.

The authors also report that the Self‑Critiquing mechanism improves the model’s self‑evaluation, enabling it to compare intermediate states and choose the more promising one, which was crucial for achieving a perfect IMO score.

Two new evaluation suites were released:

VibeSearchBench assesses whether an agent can infer user intent through multi‑turn clarification across 20 domains and 200 tasks.

VibeLifeBench tests robustness to changing external conditions over 10 domains, 20 tasks, each with 20‑30 stages and 1 247 atomic checks.

Results show that even top‑tier models like Claude Opus 5 and GPT‑5.5 fail to meet the passing threshold on these benchmarks, while dots3‑note demonstrates superior performance, especially in the later stages of long‑duration tasks.

Future work will explore recursive self‑critiquing, allowing the model to judge its progress without clear external rewards, and expand the model family (jazz, aria) to cover varying task complexities, response speeds, and compute costs.