# dots3-note Preview

> Source: https://aiwiki.ai/wiki/dots3_note
> Updated: 2026-08-15
> Categories: AI Models, Large Language Models, Mixture of Experts, Multimodal AI, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "dots3-note Preview." aiwiki.ai, 15 Aug 2026. https://aiwiki.ai/wiki/dots3_note
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**dots3-note Preview** is an open-weight [multimodal model](https://aiwiki.ai/wiki/multimodal_model) and [large language model](https://aiwiki.ai/wiki/large_language_model) developed by dots studio, an AI team at Xiaohongshu. Chao Qiao announced the preview on August 13, 2026, and the release repositories identify Xiaohongshu as the copyright holder.[1][2] The model accepts text, images, video, and audio as input and generates text output.[2]

The language backbone is a [Mixture of Experts (MoE)](https://aiwiki.ai/wiki/mixture_of_experts) model with 280 billion total parameters and 16 billion activated parameters per token. Separate encoders add visual and audio processing: the MoE vision encoder has 7 billion total parameters with 1.2 billion activated, while the dense audio encoder has 800 million parameters.[2] dots studio described it as the first open-weight release in the dots3 family and the family's most lightweight member.[2]

The word **Preview** is part of the release name and describes its maturity. At launch, dots studio provided weights, configuration files, benchmark appendices, and deployment guidance, but the model card still labeled the full report as "coming soon." No published report or peer-reviewed paper about dots3-note Preview was available when checked on August 14, 2026.[2][6]

## Release and artifacts

The release has two official checkpoint variants. Both are public and ungated on [Hugging Face](https://aiwiki.ai/wiki/hugging_face), with ModelScope mirrors linked from the model card.[2][3][5]

| Variant | Official identifier | Precision | License |
|---|---|---|---|
| Full checkpoint | `dots-studio/dots3-note-prev` | BF16 | [Apache License 2.0](https://aiwiki.ai/wiki/apache_license) |
| Quantized checkpoint | `dots-studio/dots3-note-prev-fp8` | Native FP8 with some BF16 tensors | Apache License 2.0 |

The BF16 repository contains 131 numbered Safetensors shards, separate vision and audio weight files, a weight index, tokenizer files, a chat template, and model and preprocessor configuration files.[3] The FP8 repository uses the same `Dots3NoteForCausalLM` architecture and declares `fp8` as its quantization method.[5] The GitHub repository contains release documentation, benchmark assets, and the Apache 2.0 license.[6]

Calling the release open-weight is more precise than treating it as a fully documented open-source training project. The public artifacts permit local inference and modification under Apache 2.0, but the release does not include the training data, complete training recipe, or a published technical report.[2][6]

## Architecture

The official model card and configuration expose the following component-level specifications.[2][4]

| Component | Published specification |
|---|---|
| Language backbone | 280B total parameters, 16B activated |
| Decoder layers | 1 dense layer and 45 MoE layers |
| Hidden size | 5,120 |
| Feed-forward size | 13,824 in the dense layer; 1,536 per expert |
| Experts | 256 routed and 1 shared; top 8 routed experts per token |
| Attention layers | 13 DSA and 33 sliding-window attention layers |
| DSA selection | Top 2,048 positions |
| Context configuration | 524,288 tokens, presented as 512K |
| Vocabulary | 152,064 tokens |
| Multi-token prediction | 1 shared layer, 1.13B parameters |
| Vision encoder | MoE [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer), 7B total and 1.2B activated |
| Audio encoder | Dense, 800M parameters |

The language stack alternates full-attention layers that use [DeepSeek Sparse Attention](https://aiwiki.ai/wiki/deepseek_sparse_attention) with dense sliding-window attention layers. DeepSeek introduced DSA as an indexed sparse-attention design for long contexts in DeepSeek-V3.2.[13] The merged vLLM implementation describes dots3-note's DSA layers as structurally related to that design but not directly interchangeable with the complete DeepSeek implementation. dots3-note combines DSA with sliding-window layers that have different attention geometry and key-value cache widths, then adds its own vision and audio towers.[10]

The model card gives parameter counts separately for the language, vision, and audio components. It does not publish a single aggregate activated-parameter count for every modality, so adding 16B, 1.2B, and 800M would imply a runtime behavior the documentation does not state. The 280B/16B label should therefore be read as the published language-backbone specification.[2]

## Modalities and generation

The same checkpoint handles text, image, video, and audio requests. Public examples use an OpenAI-compatible chat-completions interface. For video, the documented processor can include the audio track when one is available.[2] The configuration specifies 16 kHz audio input and a Whisper-derived encoder configuration, while image and video data pass through the MoE vision encoder.[4][10]

The chat template supports thinking and direct-response modes through an `enable_thinking` option. It also defines an XML wrapper for tool calls. vLLM's optional `dots` parser can decode that format, including multiple calls in one response, but automatic tool selection must be enabled by the server operator.[2][10]

The configured context maximum is 524,288 tokens. This is not the same as a guaranteed context length for every deployment because multimodal inputs, output reservation, concurrency, and key-value cache memory all consume the serving budget. The release does not state a general maximum output length. Its generation configuration only supplies end-of-sequence token identifiers.[4]

## Deployment

[vLLM](https://aiwiki.ai/wiki/vllm) merged native dots3-note multimodal support into its main branch on August 12, one day before the public announcement. The implementation covers text, image, audio, and native video requests, FP8 MoE inference, optional multi-token speculative decoding, and the `dots` tool-call parser. Validation used the FP8 checkpoint on an eight-GPU Hopper system.[10]

The official vLLM recipe requires eight NVIDIA GPUs with at least 80 GB each for the native FP8 checkpoint. It reports about 576.9 GB of indexed BF16 weights and budgets 692 GB of aggregate VRAM for its BF16 example.[9] The recipe uses a validated serving limit of 262,144 tokens with an 8,192-token chunked-prefill budget, even though the model configuration allows 524,288 tokens. It advises increasing the serving limit only after checking available key-value cache memory.[9]

Runtime support remained uneven at launch. Native vLLM code was available on its main branch and through nightly builds, not yet through a named stable release cited by the model card.[2][9] The [Hugging Face Transformers](https://aiwiki.ai/wiki/transformers_library) integration was still an open pull request on August 14. That pull request targeted inference and reported successful BF16 and FP8 tests for all four input modalities, but it had not entered the main branch.[11] The [SGLang](https://aiwiki.ai/wiki/sglang) integration was also open, so the official instructions pointed users to a development container or the pull-request revision.[2][12]

## Vendor-reported evaluation

The release includes reasoning, agentic, coding, search, vision, video, document, and audio [benchmark](https://aiwiki.ai/wiki/benchmark) tables. These are vendor-reported evaluations rather than independent reproductions. The tables combine official leaderboard results, dots studio's own tests, and comparator runs using different harnesses and settings.[7][8]

| Benchmark | Reported dots3-note Preview score | Published qualification |
|---|---:|---|
| ARC-AGI-2 | 81.4 | Vendor run on the official public evaluation set |
| Terminal-Bench 2.1 | 75.1 | Terminus-2 harness, 10-hour timeout, 256K context |
| ClawEval, Pass^3 | 73.4 | General set of 199 tasks, OpenClaw-style system prompt |
| VibeLifeBench | 28.1 | `avg@3`, version 1.0.0, OpenClaw harness with 256K context |
| MMMU Pro | 79.1 | Multimodal benchmark table |
| PerceptionBench | 53.4 | Multimodal benchmark table |
| MME Video-V2 | 39.3 | Audio input enabled for dots3-note and the listed Gemini comparison |
| MMAU-Pro | 69.8 | Audio-understanding benchmark table |

The ARC-AGI-2 figure has the clearest caveat. dots studio evaluated the public set with temperature 0.7, top-p 1.0, and a maximum of 384K tokens. The appendix distinguishes that result from unmarked official scores on the private leaderboard, so 81.4 should not be described as an official private-set leaderboard score.[7]

Terminal-Bench 2.1 used JSON parsing, temperature 0.7, top-p 0.95, a maximum of 81,920 tokens, and eight CPUs plus 16 GB of RAM per task. ClawEval used temperature 1.0, top-p 0.8, a maximum output length of 16,384 tokens, and a 262K context window across 38 multi-turn and 161 single-turn tasks.[7] These settings show why scores from different rows or vendors are not automatically comparable.

Some search and multimodal evaluations also depend on model-based judging or answer extraction. The appendix names GPT-OSS-120B, Qwen3.5-397B-A17B, Gemini 3.5 Flash, and other models for different judging roles.[7][8] The resulting table documents dots studio's evaluation program, but it does not establish an independently verified rank across all listed systems.

## TEMPO companion announcement

Qiao's launch post introduced TEMPO alongside dots3-note Preview. The acronym expands to **Test-time-scaled Value Estimation with Macro-step Policy Optimization**. Qiao described it as a [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) method that turns intermediate self-critique into learning signals before a long-running agent rollout finishes.[1]

The available evidence does not show that dots3-note Preview was trained with TEMPO. The announcement places TEMPO "beyond the model," while the model card does not mention the method, link a paper, or provide an implementation.[1][2] Until the promised report or another primary source states the relationship, TEMPO is best treated as a companion research announcement rather than a verified part of this checkpoint's training history.

## Documentation and safety gaps

The preview model card contains no safety, responsible-use, red-team, or risk-evaluation section.[2] That omission does not establish that the model lacks safety training. It means the release does not provide evidence that would support claims about its safety evaluation, safeguards, or resistance to misuse.

The public documentation also does not identify the training corpus, data provenance and licenses, filtering process, knowledge cutoff, training compute, energy use, post-training mixture, maximum output length, or model-level misuse mitigations.[2][4][6] Apache 2.0 grants broad permissions and disclaims warranty, but the license is not a safety assessment or a production-readiness statement.

The pending full report, open Transformers and SGLang integrations, large hardware requirement, and lack of independent replication are consistent with the preview label. They do not make the published architecture or weights unreal, but they limit stronger claims. At launch, the evidence supports describing dots3-note Preview as a real, runnable open-weight model. It does not support describing it as production stable, peer reviewed, independently benchmarked, or fully documented.

## References

1. Chao Qiao. *dots3-note Preview announcement*. August 13, 2026. https://x.com/ChaoQiao42/status/2088004290556436991
2. dots studio. *dots3-note Preview model card*. https://huggingface.co/dots-studio/dots3-note-prev
3. Hugging Face Hub API. *dots-studio/dots3-note-prev repository metadata*. Accessed August 14, 2026. https://huggingface.co/api/models/dots-studio/dots3-note-prev
4. dots studio. *dots3-note Preview config.json*. https://huggingface.co/dots-studio/dots3-note-prev/blob/main/config.json
5. dots studio. *dots3-note Preview FP8 model repository*. https://huggingface.co/dots-studio/dots3-note-prev-fp8
6. dots studio. *dots3-note Preview GitHub repository and Apache 2.0 license*. https://github.com/studio-dots-ai/dots3-note-prev
7. dots studio. *Reasoning and Agentic Evaluation appendix*. https://huggingface.co/dots-studio/dots3-note-prev/blob/main/assets/benchmark_appendix_en_reasoning.png
8. dots studio. *Multimodal Evaluation appendix*. https://huggingface.co/dots-studio/dots3-note-prev/blob/main/assets/benchmark_appendix_en_multimodal.png
9. vLLM. *dots-studio/dots3-note-prev deployment recipe*. Updated August 14, 2026. https://recipes.vllm.ai/dots-studio/dots3-note-prev
10. vLLM. *Add native Dots3 NOTE multimodal support*. Pull request 51255, merged August 12, 2026. https://github.com/vllm-project/vllm/pull/51255
11. Hugging Face Transformers. *Add dots3-note Preview model support*. Pull request 47844. https://github.com/huggingface/transformers/pull/47844
12. SGLang. *Complete dots.note.omni support with native encoders, video preprocessing, and MTP decoding*. Pull request 33829. https://github.com/sgl-project/sglang/pull/33829
13. DeepSeek-AI et al. *DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models*. arXiv:2512.02556, December 2, 2025. https://arxiv.org/abs/2512.02556