Overview
On August 14 at 7:25 PM (Shanghai time), Xiaohongshu's dots model lab published the weights of dots3-note Preview on Hugging Face and GitHub. This is the first open-source release of the dots3 series, and the first batch of community-facing results from the same model family that scored a officially certified 42/42 perfect gold medal at IMO 2026.
Key specifications:
- 280B total parameters, 16B activated parameters (MoE)
- 512K context length
- Tri-modal understanding: text, vision, speech
- Apache 2.0 license; model architecture submitted to Transformers
- No unique correct answer
- May span hours to days
- Constraints revealed incrementally
- Mid-process changes such as price shifts and weather
- The task is split into multiple macro-steps, each containing multi-turn model-environment interaction
- At the end of each macro-step, the same agent switches from actor to critic, using test-time-scaled inference to estimate expected remaining return
- This makes "evaluation easier than generation" actually hold for long-horizon tasks
- VibeSearchBench: 20 domains, 200 tasks; tools limited to search / visit / code
- VibeLifeBench: 10 domains, 20 cross-phase tasks, each spanning 20–30 stages with 1,247 atomic checks evaluating cross-stage state consistency, tool execution results, and final deliverables
- IMO 42/42 proves the ceiling of reasoning ability; open-sourcing dots3-note Preview shifts the competition toward deployment-relevant capability. With 70% of Codex users already delegating hour-plus work, crossing the "hours to days" threshold depends on cross-phase self-evaluation, long-horizon credit assignment, and real-world constraint handling — not single-shot reasoning.
- For China's domestic compute stack, the same-day 0-Day Ascend adaptation is notable: vLLM Ascend support arrived in the same release window rather than months of piecemeal community patches, which matters for sovereign/controlled deployment scenarios.
- A technical report is promised within one week, disclosing training data volume, compute consumption, and detailed head-to-head comparisons.
- Honestly acknowledged limitations: insufficient RL training, hallucination suppression, text-vs-multimodal capability balance, and stability — all pointing to the need for harnesses, connectors, data sources, and safety/permission mechanisms before real-life tasks become reliable real-world experiences.
What Was Released
The open-source release bundles three components built around one idea: solving long-horizon tasks requires an evaluation foundation plus a training paradigm.
1. dots3-note Preview weights 2. Two real-task benchmark environments: VibeSearchBench and VibeLifeBench 3. TEMPO training framework — Test-time-scaled Value Estimation with Macro-step Policy Optimization
Huawei responded the same day: Ascend Atlas 800 A3 / 900 A3 SuperPoD completed 0-Day full adaptation of the vLLM Ascend engine, boosting MoE inference throughput via FlashComm communication optimization, FUSED_MC2 operator refactoring, and MTP speculative decoding.
The Problem It Targets
Modern LLMs excel at closed-loop tasks with clear right/wrong answers (math, coding, engineering). But real-life tasks like trip planning, home renovation, or wedding organization share several difficulties:
Xiaohongshu's goal: push the model from "math competition contestant" to "executor of long-horizon real-life tasks."
TEMPO Methodology
TEMPO is a reinforcement learning paradigm that does not rely on terminal sparse rewards:
Reported results on ARC-AGI-3: TEMPO averages 31.5% above the baseline checkpoint and 20.6% above GRPO.
Benchmarks
Unlike mainstream closed-book benchmarks, current top domestic and international models fail to cross the passing line on these benchmarks.