Tasks and inputs
Hy-Embodied-RxBrain-1.0 is listed for Any-to-any, Reinforcement learning. It accepts text, image, video and produces text, image. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Hy-Embodied-RxBrain-1.0 is Tencent's 6.2B embodied model that reasons in text and imagines goal images, for planning, world-state prediction and video QA.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under Apache 2.0; no hosted price.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Most robot planners describe the next step in words. Hy-Embodied-RxBrain-1.0, which Tencent calls RxBrain for short, also draws it. Within one autoregressive sequence the model alternates reasoning text with imagined frames, and a learned token decides when it is time to picture the outcome. A plan therefore pairs what to do with what the scene should look like afterwards.
It comes from Tencent Robotics X, Futian Laboratory and the Tencent Hy team, released in July 2026 with weights, inference code and a technical report.
| Component | Detail |
|---|---|
| Repository | tencent/Hy-Embodied-RxBrain-1.0 |
| Backbone | ~6.2B-parameter Mixture-of-Transformers with separate text, vision and generation pathways |
| Image head | flow matching into a frozen FLUX VAE latent space |
| Extra file | FLUX VAE ae.safetensors (83.8M), obtained separately |
| Licence | Apache 2.0 |
The VAE dependency splits usage in two. Visual question answering needs only the main weights. Anything that produces pictures, from text-to-image to world-model rollout and interleaved planning, also needs the FLUX autoencoder.
This is research code, and it shows. A stock Transformers release does not include the hunyuan_vl_mot backbone, so you must install a pinned commit. PyTorch 2.10 and CUDA 12.x are required. The model uses its own interleaved decoding loop rather than model.generate, and the loader wants a local directory, not a Hub ID. Tencent's to-do list still includes a public RxBrain-Bench and fine-tuning code for VQA, multi-frame generation and interleaved generation, so you cannot yet adapt it with official tooling.
Hy-Embodied-VLM-1.0 shipped the same week and answers in text only, at about 30B total parameters with 3B active. RxBrain is smaller and trades raw benchmark breadth for visual imagination. Pick the VLM when you need fast, text-out decisions; pick RxBrain when a goal image is useful downstream, for example as a target for a low-level policy.
A quick feasibility test is to feed it a short manipulation clip, ask for a three-step plan, and check whether each imagined goal frame is physically plausible and consistent with the previous one. Errors there propagate straight into control. Other Tencent work sits on the Tencent provider page, and similar unified models appear under any-to-any.
Compare the model on the work you need to complete.
Hy-Embodied-RxBrain-1.0 is listed for Any-to-any, Reinforcement learning. It accepts text, image, video and produces text, image. Use the publisher’s input format and preprocessing for your chosen deployment.
Choose the output size and quality your design needs. Include drafts and revisions when budgeting an accepted image.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Liquid AI
LFM3-Omni-2B is Liquid AI's open any-to-any model: image, video, audio and text in; text and speech out, with multimodal tool calls on-device.
NVIDIA
NLD-Uni (Nemotron-Labs-Diffusion-Uni-8B) is NVIDIA's open diffusion-language model that both understands images and generates them, for research.
NVIDIA
PixelUMM is NVIDIA's encoder-free 15B unified model that understands and generates text, images and video directly in pixel space. Non-commercial.
shanghai-ai-lab
InternLumina-U2 is Shanghai AI Lab's masked-diffusion MoE that understands images, video and 3D and generates and edits images with 8-codebook visual tokens.
sensetime
SenseNova-U1.5-8B-MoT is SenseTime's open unified model for native 4K image generation and instruction-based image editing, Apache-2.0 licensed.
google/gemma-4-E2B-it is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.