Alibaba-DAMO-Academy/RynnBrain1.1-9B

RynnBrain 1.1

Towards More Capable and Generalizable Embodied Foundation Models

💫 Project Page   |   🤗 Hugging Face   |   🤖 ModelScope   |   📚 Cookbooks   |   📄 arXiv

News

• [2026.07.16] ✨✨ Release RynnBrain 1.1 checkpoints (2B / 9B / 122B-A10B) on Hugging Face & ModelScope!
• [2026.07.16] ✨✨ Release the RynnBrain 1.1 Technical Report on arXiv!

Introduction

We present RynnBrain 1.1, a systematic upgrade of RynnBrain for embodied intelligence. RynnBrain 1.1 is released in three scales: 2B, 9B, and 122B-A10B, extending the model family from compact dense models to its first 122B-level sparse-MoE model.

What's New in 1.1 🚀

• Unified Embodied Scaling to 122B: Establishes the first embodied brain model at the 122B scale under a unified training recipe shared across 2B, 9B, and 122B-A10B, enabling a systematic study of how embodied cognition, spatial reasoning, grounding, and planning evolve with scale.
• Real-robot VLA transfer: Bridges perception and action through RynnBrain-VLA, translating embodied understanding into real-robot control and demonstrating strong cross-platform generalization on Unitree G1, Astribot, and Tianji-Wuji across humanoid, bimanual, and dexterous-hand tasks.
• Native 3D and contact point grounding: Introduces explicit 3D-grounded training and a new contact point prediction task, extending RynnBrain from image-plane localization to metric 3D understanding and action-relevant interaction grounding.

Performance

• General Embodied Understanding

RynnBrain 1.1-9B vs. other 9B-scale models

• Real-Robot VLA Evaluation

Real-robot VLA success rates vs. baselines

• 3D Grounding

3D grounding accuracy vs. baselines

Model Zoo

Model	Base Model	HuggingFace	ModelScope
RynnBrain1.1-2B	Qwen3.5-2B	Link	Link
RynnBrain1.1-9B （This checkpoint）	Qwen3.5-9B	Link	Link
RynnBrain1.1-122B-A10B	Qwen3.5-122B-A10B	Link	Link

Quick Start

Inference with 🤗 transformers

Minimal dependencies

pip install transformers==5.2.0

Run text generation

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "cookbooks/assets/object_location/images/000000086408.jpg"},
            {
                "type": "text",
                "text": "What appliance can be used to heat food quickly.\nGenerate coordinates for one object bounding box. Constraints: x1,y1,x2,y2 in [0,1000]. Response must be in the format: <object> (x1, y1), (x2, y2) </object>",
            },
        ],
    }
]

model_path = "Alibaba-DAMO-Academy/RynnBrain1.1-9B"
processor = AutoProcessor.from_pretrained(model_path)

model = AutoModelForImageTextToText.from_pretrained(
    model_path,
    dtype=torch.bfloat16,
)
model.to("cuda")

model_inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    enable_thinking=False,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
)
model_inputs = model_inputs.to("cuda")

output_ids = model.generate(
    **model_inputs,
    max_new_tokens=256,
    do_sample=False,
)
output_ids = output_ids[:, model_inputs["input_ids"].size(1) :]
response = processor.decode(output_ids[0], skip_special_tokens=True)
print(response)

Inference with SGLang

For installation and advanced usages, please refer to the official documentation.

OpenAI-Compatible Serving

# launch server
python3 -m sglang.launch_server --model-path Alibaba-DAMO-Academy/RynnBrain1.1-9B --host 0.0.0.0 --port 8000

# inference using openai api
import base64
import io

from openai import OpenAI
from PIL import Image

def pil_to_url(image: Image.Image):
    image_format = image.format if image.format else 'PNG'
    buffered = io.BytesIO()
    image.save(buffered, format=image_format)
    img_str = base64.b64encode(buffered.getvalue()).decode('utf-8')
    return f'data:image/{image_format.lower()};base64,{img_str}'

messages = [
    {
        'role': 'user',
        'content': [
            {'type': 'image_url', 'image_url': {'url': pil_to_url(Image.open('cookbooks/assets/object_location/images/000000086408.jpg'))}},
            {'type': 'text', 'text': 'What appliance can be used to heat food quickly.\nGenerate coordinates for one object bounding box. Constraints: x1,y1,x2,y2 ∈ [0,1000]. Response must be in the format: <object> (x1, y1), (x2, y2) </object>'},
        ],
    }
]

client = OpenAI(api_key="", base_url="http://localhost:8000/v1")
response = client.chat.completions.create(
    model="default",
    messages=messages,
    stream=False,
).choices[0].message.content
print(response)

Offline Engine

import sglang as sgl
from transformers import AutoProcessor

def main():
    conversation = [
        {
            'role': 'user',
            'content': [
                {'type': 'image'},
                {'type': 'text', 'text': 'What appliance can be used to heat food quickly.\nGenerate coordinates for one object bounding box. Constraints: x1,y1,x2,y2 ∈ [0,1000]. Response must be in the format: <object> (x1, y1), (x2, y2) </object>'},
            ],
        }
    ]

    model_path = 'Alibaba-DAMO-Academy/RynnBrain1.1-9B'
    llm = sgl.Engine(model_path=model_path)
    processor = AutoProcessor.from_pretrained(model_path)

    prompt = processor.apply_chat_template(
        conversation,
        add_generation_prompt=True,
        tokenize=False,
    )

    output = llm.generate(
        prompt=prompt,
        image_data='cookbooks/assets/object_location/images/000000086408.jpg',
        sampling_params={"temperature": 0.8, "top_p": 0.95},
    )
    print(f"Prompt: {prompt}\nGenerated text: {output['text']}")

if __name__ == '__main__':
    main()

Cookbooks

Check out the cookbooks that showcase RynnBrain's capabilities in cognition, localization, reasoning, and planning.

Category	Cookbook name	Description
Spatial Understanding	1_spatial_understanding.ipynb	Shows the model's ability for spatial understanding in the video scene.
Object Understanding	2_object_understanding.ipynb	Shows how the model understands object categories, attributes, and relations and counting ability.
Object Grounding	3_object_grounding.ipynb	Locates specific objects with bounding boxes in an image or video based on instructions.
Area Location	4_area_location.ipynb	Identifies and marks specified regions by points in an image or video.
Affordance Location	5_affordance_location.ipynb	Finds areas or objects with specific affordances in an image or video.
Trajectory Location	6_trajectory_location.ipynb	Infers and annotates trajectories or motion paths in an image or video.
🆕 Contact Point Prediction	7_contact_point_prediction.ipynb	Predicts an instruction-conditioned contact point and in-plane orientation from an image.
🆕 3D Grounding	8_3d_grounding.ipynb	Predicts 3D bounding boxes (position, dimensions, orientation) from a single RGB image with camera intrinsics.

Training

Pretraining & Evaluation

Please refer to RynnScale for details of pretraining and evaluation.

📑 Citation

If you find RynnBrain useful for your research and applications, please cite using this BibTeX:

@article{damo2026rynnbrain,
  title={RynnBrain: Open Embodied Foundation Models},
  author={Ronghao Dang, Jiayan Guo, Bohan Hou, Sicong Leng, Kehan Li, Xin Li, Jiangpin Liu, Yunxuan Mao, Zhikai Wang, Yuqian Yuan, Minghao Zhu, Xiao Lin, Yang Bai, Qian Jiang, Yaxi Zhao, Minghua Zeng, Junlong Gao, Yuming Jiang, Jun Cen, Siteng Huang, Liuyi Wang, Wenqiao Zhang, Chengju Liu, Jianfei Yang, Shijian Lu, Deli Zhao},
  journal={arXiv preprint arXiv:2602.14979v1},
  year={2026},
  url = {https://arxiv.org/abs/2602.14979v1}
}