Model Details
dots3-note Preview
Quick answer: dots3-note Preview is the first open-weight model in Dots Studio's new dots3 family — a 280B-total/16B-activated Mixture-of-Experts model built by Dots Studio, the AI lab behind Xiaohongshu (RedNote), supporting text, image, video, and audio input with text output and up to 512K tokens of context. Released under Apache 2.0, it scores 78.4% on SWE-bench Verified, 61% on SWE-bench Pro, and 75.1% on Terminal Bench 2.1.
At a Glance
Where dots3-note Preview leads
- 78.4% SWE-bench Verified — strong real-world coding agent performance
- 75.7% SWE-bench Multilingual — solid cross-language coding generalization
- 79.1% MMMU Pro (standard, 10 options) — strong multimodal reasoning for its lightweight-tier positioning
- Native text, image, video, and audio understanding in a single checkpoint (dedicated 800M audio encoder + 7B/1.2B-active MoE vision encoder)
- Apache 2.0 — fully open weights
Where it lags
- 61% SWE-bench Pro — noticeably behind SWE-bench Verified, consistent with SWE-bench Pro's harder, more adversarial task design across all models
- 30.8% Apex Agents — modest on long-horizon agentic tool-use tasks
- 52.8% SkillsBench v1.1 — mid-pack on applied-skills evaluation
Best for: self-hosted multimodal agent workloads (coding, tool use, document/image/video/audio understanding) where a single checkpoint needs to handle multiple modalities without separate encoders bolted on after the fact.
What dots3-note Preview Is
dots3-note Preview is the first open-weight release in Dots Studio's dots3 model family — Dots Studio being the AI research team behind Xiaohongshu (RedNote), the Chinese social platform. It's a Mixture-of-Experts model with 280B total parameters and 16B activated parameters per forward pass, positioned by Dots Studio as the lightest-weight member of a planned family of models with different capability/latency/cost trade-offs.
Architecturally, dots3-note Preview pairs a 45-layer MoE language backbone (256 routed + 1 shared expert, top-8 routing, plus 1 dense layer) with a dedicated MoE vision encoder (7B total / 1.2B activated) and a dense 800M-parameter audio encoder — giving it native understanding of text, images, video, and audio in a single model, with text-only output. Context length extends to 512K tokens. The model supports both BF16 and FP8 precision, with FP8 the recommended path for production serving (one 8-GPU node via SGLang or vLLM).
Aimed at general knowledge and instruction following, mathematical/logical reasoning, tool use and multi-step agent workflows, and code generation, dots3-note Preview reports competitive scores on real-world coding benchmarks (SWE-bench Verified 78.4%, SWE-bench Pro 61%, Terminal Bench 2.1 75.1%) alongside solid multimodal understanding (MMMU Pro 79.1%). It trails on longer-horizon agentic evaluations like Apex Agents (30.8%) and SkillsBench v1.1 (52.8%), suggesting its agent-specific training is less mature than its coding/multimodal foundations at this preview stage.
Specs
| Field | Value |
|---|---|
| Organization | Dots Studio (Xiaohongshu) |
| Architecture | Multimodal Mixture-of-Experts |
| Total parameters | 280B |
| Activated parameters | 16B (+ 1.13B shared MTP layer) |
| Layers | 1 dense + 45 MoE |
| Experts | 256 routed + 1 shared, top-8 routing |
| Vision encoder | MoE ViT, 7B total / 1.2B activated |
| Audio encoder | Dense, 800M |
| Context window | 512,000 tokens |
| Input modalities | Text, image, video, audio |
| Output modality | Text |
| Precision | BF16, FP8 |
| License | Apache 2.0 |
| Release date | August 2026 |
Pricing
dots3-note Preview is released as open weights (self-hosted); Dots Studio has not announced a hosted API pricing tier for this preview model. It's available free to try via OpenRouter.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| SWE-bench Verified | 78.4% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| SWE-bench Pro | 61% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| Terminal Bench 2.1 | 75.1% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| WildClawBench | 61.7% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| MMMU Pro | 79.1% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| Apex Agents | 30.8% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
| SkillsBench v1.1 | 52.8% | dots3-note-prev model card (HF Eval Results) | 2026-08 |
Scores are self-reported by Dots Studio on the official Hugging Face model card (dots-studio/dots3-note-prev), sourced directly from the card's structured HF "Evaluation Results" metadata section (the most reliable source, since it unambiguously ties each score to this specific model rather than requiring interpretation of a prose comparison table). Terminal Bench 2.1, Apex Agents, and SkillsBench v1.1 carry an asterisk on the model card indicating a caveat in the source methodology notes — treat as directionally accurate but confirm against Dots Studio's benchmark appendix if precision matters.
Notable Results Not Yet Tracked as Benchgen Benchmarks
- SWE-bench Multilingual (75.7%) is a real, independently-documented HF dataset (
SWE-bench/SWE-bench_Multilingual) confirmed via the model card's structured Eval Results, but doesn't have a standalone Benchgen benchmark page yet — flagged as a follow-up. - The model card also references broader "General Reasoning and Agent" and "Multimodal Understanding" benchmark suites via image-only appendix charts (not machine-readable prose/tables) — not included here pending a text-based source.
FAQ
Is dots3-note Preview open source?
Yes — Apache 2.0, with weights available on Hugging Face (dots-studio/dots3-note-prev, plus an FP8-quantized variant).
Who makes dots3-note Preview? Dots Studio, the AI research team at Xiaohongshu (RedNote), the Chinese social media platform.
How big is dots3-note Preview? 280B total parameters, 16B activated per forward pass (Mixture-of-Experts), plus a 7B/1.2B-active MoE vision encoder and an 800M dense audio encoder.
What modalities does dots3-note Preview support? Text, image, video, and audio as input; text as output.
What is dots3-note Preview's context window? 512,000 tokens.
How does dots3-note Preview perform on coding benchmarks? 78.4% on SWE-bench Verified, 61% on SWE-bench Pro, and 75.1% on Terminal Bench 2.1 — strong for a preview-stage, "lightest" member of its model family.
Where can I access dots3-note Preview? Self-hosted from the open weights at huggingface.co/dots-studio/dots3-note-prev (BF16 or FP8), via Transformers, SGLang, or vLLM. Also available free via OpenRouter.
Benchmark scores sourced from the official Hugging Face model card for dots-studio/dots3-note-prev (Aug 2026), taken directly from the card's structured HF Eval Results metadata.