dots-studio / dots3-note-prev-fp8

中文 ｜ English

🌐 Tech Blog |   📄 Full Report (coming soon)

---

Table of Contents

• Model Introduction
• Model Overview
• Evaluation Results
  • General Reasoning and Agent
  • Multimodal Understanding
• Model Links
• Quickstart
• Deployment
  • Transformers
  • SGLang
  • vLLM
• Benchmark Appendix
• License
• Contact Us

---

Model Introduction

dots3-note preview is the first open-weight model in the dots3 family. It is a Mixture-of-Experts model with 280B total parameters, 16B activated parameters, and support for a context length of up to 512K tokens. The model can understand text, images, video, and audio, and produces text outputs.

dots3-note preview is optimized for a broad range of tasks, including:

• General knowledge and instruction following;
• Mathematical and logical reasoning;
• Tool use and multi-step agent workflows;
• Interactive tasks that require exploration, memory updates, and adaptation;
• Code generation and code-based problem solving;
• Image, document, chart, audio, and video understanding;
• Long-context information processing.

The dots3 family is designed to include models with different trade-offs among capability, latency, and inference cost. dots3-note preview is the most lightweight member of the family.

Model Overview

Evaluation Results

General Reasoning and Agent

Multimodal Understanding

Model Links

Quickstart

Recommended: serve the FP8 checkpoint on one 8-GPU node with SGLang or vLLM .

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="dots3-note-prev",
    messages=[
        {"role": "user", "content": "Hello! Can you briefly introduce yourself?"},
    ],
    temperature=1.0,
    top_p=0.95,
    max_tokens=256,
    # Set enable_thinking=True for reasoning; False returns a direct response.
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)

For a multimodal request, replace messages with one of these public examples:

examples = {
    "image": [
        {"type": "image_url", "image_url": {"url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png"}},
        {"type": "text", "text": "How many cats are in this image?"},
    ],
    "audio": [
        {"type": "audio_url", "audio_url": {"url": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/mary_had_lamb.mp3"}},
        {"type": "text", "text": "Transcribe this nursery rhyme."},
    ],
    "video": [
        {"type": "video_url", "video_url": {"url": "https://huggingface.co/datasets/merve/vlm_test_images/resolve/main/concert.mp4"}},
        {"type": "text", "text": "Describe the performance and what can be heard."},
    ],
}
messages = [{"role": "user", "content": examples["image"]}]

Video inputs include their audio track when available.

Deployment

The commands below target FP8 on one 8-GPU node. BF16 requires more memory. Tune the context length to available memory, concurrency, and input modalities.

Native support is available on vLLM main . Transformers #47844 and SGLang #33829 are still under review; until they are merged, use the PR revisions below.

Transformers

First install mutually compatible PyTorch and torchvision builds supported by your NVIDIA driver. For audio and video, also install a PyTorch-compatible torchcodec (included below) and FFmpeg with your system package manager. Then install Transformers #47844 :

pip install accelerate pillow torchcodec kernels==0.16.0 "transformers @ git+https://github.com/huggingface/transformers.git@refs/pull/47844/head"

Run a minimal local inference:

from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "dots-studio/dots3-note-prev-fp8"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(model_id, dtype="auto", device_map="auto")

messages = [
    {"role": "user", "content": "Hello! Please briefly introduce yourself."},
]
inputs = processor.tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    enable_thinking=False,
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(outputs[0, inputs.input_ids.shape[1] :], skip_special_tokens=True))

Use SGLang or vLLM for multi-GPU OpenAI-compatible serving.

SGLang

Recommended: use the release image lmsysorg/sglang:dev-dots3-note . Full one-node recipes and tuning notes are in the Dots3-Note cookbook . Source support is tracked in SGLang #33829 .

Docker (the image downloads the checkpoint from Hugging Face on first run):

docker run --gpus all --ipc=host -p 8000:8000 \
  lmsysorg/sglang:dev-dots3-note \
  sglang serve \
    --model-path dots-studio/dots3-note-prev-fp8 \
    --served-model-name dots3-note-prev \
    --host 0.0.0.0 \
    --port 8000 \
    --context-length 524288 \
    --enable-dp-attention \
    --dp-size 8 \
    --tp-size 8 \
    --ep-size 8 \
    --moe-dense-tp-size 1 \
    --page-size 64 \
    --trust-remote-code \
    --attention-backend fa3 \
    --moe-a2a-backend deepep \
    --enable-multimodal \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --speculative-draft-model-path dots-studio/dots3-note-prev-fp8

Or install from source / the PR and run the same sglang serve arguments locally. --attention-backend fa3 sets prefill, decode, and (when speculative decoding is enabled) draft attention. MTP/NEXTN ( --speculative-algorithm NEXTN and the related flags) is optional and can reduce TPOT by more than 50%. Prefill CUDA graph is not supported yet.

Optional features:

# Load only the language model
--language-only

# Enable OpenAI-compatible tool calling
--tool-call-parser dots

vLLM

Native dots3-note preview support is available on vLLM main . Use a recent nightly build until it is included in a stable release.

The following example deploys the FP8 checkpoint on eight NVIDIA H100 GPUs with TP=8 and EP=8:

vllm serve dots-studio/dots3-note-prev-fp8 \
  --served-model-name dots3-note-prev \
  --host 0.0.0.0 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --moe-backend deep_gemm \
  --max-model-len 262144

Optional features:

# Load only the language model
--language-model-only

# Enable three-token MTP speculative decoding
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'

# Enable OpenAI-compatible automatic tool calling
--enable-auto-tool-choice --tool-call-parser dots

Benchmark Appendix

License

Copyright (c) 2026 Xiaohongshu.

Developed and released by dots studio.

The dots3-note preview model weights and modeling code in this repository are licensed under the Apache License, Version 2.0.

See the LICENSE file for details.

Transformers, SGLang, vLLM, and other third-party software are subject to their respective licenses.

Contact Us

For questions and feedback, please contact us through:

• Email: [email protected]

---

dots3-note preview is developed and released by dots studio.