RynnBrain 1.1 is an embodied brain model that also points at 3D contact points

One architecture across three scales, including a sparse 122B

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

The architectural claim is that the same recipe is shared across all three. The README says the unified scaling establishes the first embodied brain model at the 122B scale under a unified training recipe shared across 2B, 9B and 122B-A10B, and the stated purpose is a systematic study of how embodied cognition, spatial reasoning, grounding and planning evolve with scale. That framing matters, because a scale study is only meaningful if the training setup is held constant, and the claim here is that it is.

Underneath, the model architecture is a unified decoder-only vision-language architecture across all scales, supporting both dense and sparse-MoE variants. The model zoo table makes the base models explicit, and they are Qwen models: RynnBrain1.1-2B is based on Qwen3.5-2B, RynnBrain1.1-9B on Qwen3.5-9B, and RynnBrain1.1-122B-A10B on Qwen3.5-122B-A10B. So the family is a set of fine-tunes of one language model family at three capacities rather than three unrelated models.

The README also states that the framework is generic enough for domains beyond agriculture, while the shipped models are optimized for agriculture and sustainability. That is a claim about the underlying tooling rather than about RynnBrain itself, and it is worth separating from the embodied claims.

Grounding moves from image plane to metric 3D and contact points

The most concrete new capability in 1.1 is in the output space. The README says the model 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.

The distinction between image-plane and metric 3D is the one to hold on to. Image-plane localization produces coordinates inside the picture, which is enough to draw a box and stop. Metric 3D grounding produces a position in space with a scale, which is what you need to answer a question like where in the room the handle is. For a robot, that difference is the difference between pointing at a pixel and reaching for a thing.

Contact point prediction is the finer-grained version of the same idea. Instead of a bounding box for an object, the model predicts where on the object contact should happen, which is the difference between knowing a mug is on the table and knowing which point of the mug to grasp.

The architecture section describes the outputs as aligned ones: text, pointing sequences, 3D perception and contact signals, all produced from omni-vision inputs combined with language instructions. A single decoder emitting several output modalities for the same scene is a different design from running a separate grounding head or a separate detector alongside the language model, and it is the reason one checkpoint can serve both a chat-style question and a coordinate request.

RynnBrain-VLA is where perception becomes robot control

The third item in the list of what is new is real-robot transfer through RynnBrain-VLA, which the README says bridges perception and action by translating embodied understanding into real-robot control.

The hardware named is specific: Unitree G1, Astribot and Tianji-Wuji, and the tasks are described as humanoid, bimanual and dexterous-hand tasks. The claim attached to it is strong cross-platform generalization on those three platforms, which is the interesting part, since a policy that only works on one robot body is a demonstration rather than a result.

The distinction between the RynnBrain 1.1 checkpoints and RynnBrain-VLA is worth noting for anyone planning an evaluation. The checkpoints are the model family, available at three scales, and RynnBrain-VLA is the layer that turns the embodied understanding into control. The repository layout does not separate the two clearly: the top-level entries are a .gitignore, the LICENSE, the README, the 1.1 technical report as a PDF, a cookbooks directory, a demo.py, an index.html and a sitemap.xml. So the visible code surface is a demo and a set of cookbooks.

The performance section is organised in four groups: general embodied understanding compared against other models at 2B, 9B and 122B scale, real-robot VLA success rates against baselines, 3D grounding accuracy against baselines, and a contact point prediction visualisation. Every one of these is an image in the cookbooks assets directory, which is the practical limit on what can be verified from the repository alone.

Running the 2B model on transformers

The quick start is a single inference path, and the dependency list is deliberately narrow. The README asks for a specific transformers version:

pip install transformers==5.2.0

The example builds a chat conversation with an image and a text instruction, loads the processor and the model, and generates a response. The model path in the example is the 2B checkpoint:

model_path = "Alibaba-DAMO-Academy/RynnBrain1.1-2B"
processor = AutoProcessor.from_pretrained(model_path)

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

The model is loaded in bfloat16 and moved to CUDA, so the example assumes a GPU. The class used is AutoModelForImageTextToText, the standard multi-modal auto class, and the processor is driven through apply_chat_template with add_generation_prompt set to true and enable_thinking set to false, so the chat template is the interface rather than a hand-built prompt string.

The instruction in the example shows the grounding output format the model is expected to produce. It asks a question about an appliance, then requests coordinates for one object bounding box with the constraint that x1, y1, x2 and y2 are in the range 0 to 1000, and specifies the response format as an object element containing the two coordinate pairs. So the coordinates are normalised to a thousand-unit grid, and the answer is expected to be a single tagged element rather than prose. That is a concrete contract you can write a parser against, and it is the contract to check first if you are building on top of it.

What the repository does and does not give you

The repository is small by design, and the reason is that most of the substance lives elsewhere. The project page, the Hugging Face collection and the ModelScope collection are all linked from the README header, as are the cookbooks, the arXiv entry and a direct link to the 1.1 technical report PDF, which is also committed at the repository root as RynnBrain_1_1.pdf.

What the repository itself holds is a demo script, an index.html, a sitemap.xml and the cookbooks directory, plus the README. The demo is therefore the reference for how to drive the model, and the cookbooks are the reference for how the authors structure a task. The documentation is thin by the standards of a library, because this is a model release and not a package with an API surface.

That has consequences for adoption. There is no training code in the top-level layout, no evaluation harness and no serving configuration, so reproducing the reported results means reimplementing the evaluation, and deploying the model means building the serving layer yourself. A team that needs either of those should treat the technical report as the primary document and the repository as a starting point rather than a foundation.

One more navigational detail: the resources for RynnBrain 1.0 are on a separate branch, rynnbrain1.0, so the main branch reflects 1.1 only. The 1.0 checkpoints and code were released on 2026.02.09 with a technical report on 2026.02.17, RynnBrain-4B was launched on 2026.04.13, and the 1.1 release with checkpoints and report landed on 2026.07.16.

Apache-2.0 terms and a branch that moved in the summer

The licence is Apache-2.0, with a LICENSE file at the repository root. For a model release, the licence question is narrower than it looks for code, because the weights are distributed on Hugging Face and ModelScope under those platforms' terms, and the model zoo table links to a separate model page for each of the three checkpoints. Anyone using the 122B-A10B variant in particular should confirm the terms on that checkpoint rather than assume the repository licence carries over.

The repository is not archived, and the last push was on 2026-08-10, a couple of months after the 1.1 release in July. There are no GitHub releases, so the checkpoints, not the repository, are the versioned artefact, and the branch is where any post-release changes would appear.

The project page and the arXiv paper are the canonical descriptions, and the README is an index to them. That distribution of information is normal for a research release from a large lab, but it changes what an adopting team has to do: read a paper, fetch a checkpoint, and write the integration. Anyone planning to build on RynnBrain 1.1 should start from the technical report and the cookbooks, since the README's own summary is a list of claims and images rather than an implementation guide.

Editorial conclusion

RynnBrain 1.1 suits a research group working on embodied perception, spatial grounding or vision-language-action control that wants checkpoints at several scales under one architecture and one training recipe, rather than assembling a separate model per capability. It is the wrong tool for a team that needs a production serving stack, because the repository ships a demo script, cookbooks and images rather than a deployment guide, and it is the wrong tool for anyone who needs the exact numbers behind the comparison charts, since every performance figure in the README is published as an image. Before you adopt it, read the 1.1 technical report rather than the README summaries, check the cookbooks directory for a recipe matching your hardware, and note that the earlier RynnBrain 1.0 resources live on a separate branch.

Frequently asked questions

Official sources

1. alibaba-damo-academy/RynnBrain on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README