English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Xiaohongshu Open-Sources dots3-note Preview: 280B MoE Model for Long-Horizon Tasks

Forum topic · 小凯 · 2026-08-17

Summary

Xiaohongshu's dots model lab released dots3-note Preview on August 14, the first open-source release in the dots3 series and the first publicly available weights from the family that achieved an official 42/42 gold score at IMO 2026. The model has 280B total parameters, 16B active parameters, 512K context length, and tri-modal understanding (text, vision, speech) under Apache 2.0, with architecture submitted to Transformers. Alongside the weights, the team open-sourced two real-task evaluation environments (VibeSearchBench covering 20 domains and 200 tasks; VibeLifeBench covering 10 domains and 20 cross-stage tasks with 1,247 atomic checks) and a training framework called TEMPO, which uses test-time-scaled value estimation with macro-step policy optimization. Reported gains include +31.5% over the baseline checkpoint and +20.6% over GRPO on ARC-AGI-3. Huawei announced same-day vLLM Ascend 0-Day full compatibility on Atlas 800 A3 / 900 A3 SuperPoD, marking a rare synchronized open-source model and domestic compute stack. Limitations noted include insufficient RL training, hallucination control, multimodal balance, and stability.

Overview

On August 14 at 7:25 PM (Shanghai time), Xiaohongshu's dots model lab released the weights of dots3-note Preview on Hugging Face and GitHub. It is the first open-source release in the dots3 series and the first publicly available weights from the model family that earned an official 42/42 gold score at the International Mathematical Olympiad (IMO) 2026.

Key specifications:

  • 280B total parameters, 16B active parameters
  • 512K context length
  • Tri-modal understanding (text, vision, speech)
  • Apache 2.0 license
  • Model architecture submitted to Transformers
  • What was released

    The open-source drop ships three components as a unified stack:

    1. dots3-note Preview weights — the model itself. 2. VibeSearchBench and VibeLifeBench — two real-task evaluation environments. VibeSearchBench covers 20 domains across 200 tasks with a search / visit / code toolset. VibeLifeBench covers 10 domains with 20 cross-stage tasks (each spanning 20–30 stages) and 1,247 atomic checks that directly assess cross-stage state consistency, tool execution results, and final deliverables. 3. TEMPO — a training framework (Test-time-scaled Value Estimation with Macro-step Policy Optimization).

    The three pieces reflect one thesis: long-horizon capability requires an evaluation substrate, a training paradigm, and a model — none alone is enough.

    The problem it targets

    Frontier models perform strongly on closed-loop, quickly-verifiable tasks (math, coding, engineering). Real-life long-horizon tasks — travel planning, renovation preparation, wedding organization — share four traits: no single correct answer, durations of hours to days, constraints that surface incrementally, and ongoing disruptions such as price and weather changes.

    The release is positioned as pushing the model from "math competitor" to "executor of long-horizon life tasks."

    TEMPO methodology

    TEMPO trains long-horizon agents without relying on terminal sparse rewards:

  • Each task is split into multiple macro-steps, with multiple model–environment turns inside each macro-step.
  • At the end of each macro-step, the same agent switches from actor to critic and uses test-time scaling to estimate the expected remaining return.
  • This design makes "evaluation is easier than generation" actually hold for long-horizon settings.

    Reported empirical results

    On ARC-AGI-3, TEMPO averaged:

  • +31.5% over the baseline checkpoint
  • +20.6% over GRPO
  • Notably, the strongest domestic and international closed models do not yet cross the pass line on VibeSearchBench and VibeLifeBench, which the authors frame as evidence that these benchmarks measure capabilities not captured by mainstream closed-form evaluations.

    Ecosystem impact

    Huawei announced same-day (0-Day) full compatibility of the vLLM Ascend engine on Atlas 800 A3 / 900 A3 SuperPoD, layering FlashComm communication optimization, FUSED_MC2 operator reconstruction, and MTP speculative decoding to push MoE inference throughput. This makes dots3-note a rare case of a domestic model paired with a domestic compute stack at the same generation and release window.

    Why it matters

    While IMO gold certifies the ceiling of single-shot reasoning, this release shifts the competitive axis toward deployment-relevant capabilities:

  • Cross-stage self-evaluation
  • Long-horizon credit assignment
  • Handling real-world constraints that arrive mid-task
  • For the "self-controlled" narrative in government and enterprise scenarios, synchronized model-and-compute availability is arguably more consequential than another SOTA number.

    Caveats and next steps

    The team promises a full technical report within one week, disclosing training data scale, training compute, and item-by-item comparisons. Stated limitations include:

  • Reinforcement-learning training is still insufficient
  • Hallucination suppression
  • Balance between text and multimodal capabilities
  • Stability
These are framed as requiring coordinated progress across harness, connectors, data sources, and safety/permission mechanisms before long-horizon life tasks become reliable real experiences.

Sources

Xiaohongshu tech WeChat account (dots.llm), lifestyle.onebyfourstudio.com, aibase.com, ModelScope community.

Tags

#long-horizon-agents#moe-model#open-source-llm#reinforcement-learning#test-time-scaling#evaluation-benchmark#ascend-compatibility#xiaohongshu

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633593