Published: 15 August 2026
Dots studio, the AI lab inside Xiaohongshu, released dots3-note Preview on 14 August 2026, its first public model in the dots3 family, under the Apache 2.0 licence. dots3-note Preview is a 280 billion parameter mixture of experts with 16 billion active parameters, a 512,000 token context window, and input across text, images, video and audio. The release introduces TEMPO, a reinforcement learning method dots studio built to train agents on tasks that take tens of hours to complete.
What can dots3-note Preview do?
Dots studio describes dots3-note Preview as a step towards long horizon agency in real life, and the design follows from that. The model takes text, images, video and audio as input and returns text, with video inputs including their audio track where one is present. It is built to explore unfamiliar environments, update memory over time and combine multimodal perception with coding and tool use across long running tasks rather than answer single questions well.
The lab shipped two evaluation environments alongside the model, VibeSearchBench and VibeLifeBench, aimed at measuring exactly that kind of extended work. Releasing the benchmark alongside the model is how a lab signals which capability it wants to be judged on.
Dots3-note Preview architecture and technical specs
The published specification is very detailed for a preview release. dots3-note Preview has 280 billion total parameters with 16 billion active, across one dense layer and 45 mixture of experts layers. It routes to 8 of 256 routed experts plus one shared expert per token, at a hidden size of 5,120, with a dense feed forward width of 13,824 and 1,536 per expert. Attention mixes 13 DSA blocks with 33 sliding window blocks, roughly one to three, with DSA operating over the top 2,048 positions. A multi token prediction layer adds a further 1.13 billion parameters.
The perception stack is separate from the language model. Vision runs through a mixture of experts vision transformer of 7 billion parameters with 1.2 billion active, and audio through a dense 800 million parameter encoder. The context window is 512,000 tokens, the vocabulary is 152,000 tokens, and the weights ship in both BF16 and FP8. Both vLLM and SGLang expose a flag to load the language model alone if the perception encoders are not needed.
On benchmarks, dots studio publishes its own table against HY3, GLM 5.2, DeepSeek V4 flash and pro, Seed 2.1 turbo, Kimi K3, Claude Opus 4.8 and GPT-5.5, and the results are:
- The model leads on abstract reasoning and search: 81.39 on ARC AGI 2 against 72.1 for Claude Opus 4.8, 60.4 for Kimi K3 and 22.8 for GLM 5.2,
- 6.9 on the official ARC AGI 3 against 1.5 for Claude Opus 4.8 and 0.4 for GPT-5.5,
- 80.4 on IFBench against 62.2 for Claude Opus 4.8.
On agentic coding it loses more often than it wins:
- 75.1 on Terminal-Bench 2.1 against 88.3 for Kimi K3 and 81 for GLM 5.2,
- 61 on SWE-bench Pro against 69.2 for Claude Opus 4.8,
- 55.6 on Toolathlon Verified against 76.5 for Kimi K3.
Dots studio publishes its harnesses and judge models, but roughly half the competitor scores in the table carry an asterisk whose meaning is never defined anywhere in the announcement.
What is TEMPO, and what did dots studio actually prove?
TEMPO stands for Test-time-scaled Value Estimation with Macro-step Policy Optimization, and it exists to solve a training problem rather than an inference one. Dots studio’s argument is that a single agent rollout on a genuinely long task can take more than ten hours, and that the reward at the end is too sparse to assign credit across it. TEMPO breaks the task into macro-steps of several interaction rounds each. At the end of a macro-step the same model switches from actor to critic and uses test-time-scaled reasoning to estimate the expected remaining return, which lets the policy update before the task finishes. In dots studio’s framing, reinforcement learning teaches the model not only how to act but how to evaluate itself.
The evidence offered for TEMPO is thin relative to the claim. Dots studio reports that TEMPO achieves an average score 31.5 percent higher than the baseline checkpoint and 20.6 percent higher than GRPO on ARC-AGI-3, but that figure appears only in a chart caption, with no table, no seed count and no variance. The full technical report covering architecture, training and evaluation was promised within a week of launch and had not been published as of 17 August 2026.
One widely repeated claim needs correcting. Dots studio states that at IMO 2026 it built an internal harness around a branch of dots3-note Preview that generated proofs recursively and used tools to evaluate and improve them, and that this system earned an officially certified perfect score of 42 out of 42 and a gold medal. That result belongs to a branch of the model inside a bespoke harness, not to the checkpoint published on 14 August. Coverage saying dots3-note scored 42 out of 42 at the IMO overstates what was released.
Where to get dots3-note Preview
The weights are on Hugging Face as dots-studio/dots3-note-prev and dots-studio/dots3-note-prev-fp8, on ModelScope under the same names, and the code is on GitHub at studio-dots-ai/dots3-note-prev. The licence is Apache 2.0, confirmed in the licence file, so commercial use is permitted. For teams that want to try it without hosting, OpenRouter carries it free at dots-studio/dots-3-note-preview with the full 512,000 token context.
Framework support is partial at launch. vLLM supports the model natively on main, while the Transformers and SGLang pull requests were still under review as of 17 August 2026, with dots studio pointing SGLang users at a dedicated development image in the meantime. First party API access is by application through the dots.ai platform.
For more information visit the official Dots studio announcement of dots-3-note Preview.
