SSyncDev

Hy-Embodied-VLM-1.0

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

Hy-Embodied-VLM-1.0 is Tencent's ~30B MoE vision-language model for robots and embodied agents, with ~3B active parameters, 32K context and hybrid reasoning.

Key specifications

Specification score
44.5
Context window
32.77K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted per-token price published.

Inputs and outputs

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

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionOpen weights

What is Hy-Embodied-VLM-1.0?

Robots cannot wait two seconds for a planner. That latency budget is the whole reason Tencent built Hy-Embodied-VLM-1.0 as a sparse model: about 30B parameters in total, only about 3B active per token. Tencent Robotics X, the Hy Vision team and Futian Laboratory pair the Hy3-A3B language backbone with the Hy-ViT2 vision encoder, routing each token through 8 of 128 experts plus a shared one.

Spec table

Repository tencent/Hy-Embodied-VLM-1.0
Release 14-15 July 2026
Parameters ~30B total, ~3B active
Context 32,768 tokens
Images up to 128 per prompt, native aspect ratio
Precision BF16 (~86 GB across GPUs)
Reasoning hybrid, switched per request
Licence Apache 2.0

Tencent's own dates differ by a day (14 July for the weights and report, 15 July in the changelog). Either works for version tracking.

The training idea

Tencent organises embodied skill into three rungs. First, understanding the state of the agent and its surroundings. Second, reasoning about what an action will change. Third, planning long sequences, reflecting, and recovering when something goes wrong. Data for both pre-training and post-training was built around those rungs. Post-training then runs a self-evolving loop of reinforcement learning and rejection-sampling fine-tuning, seeded with a small set of curated reasoning traces, and finishes by fusing separately trained continuous-reward and discrete-reward policies.

Claims worth checking

Tencent reports that the model ranks first on 19 of 38 embodied benchmarks and second on another 11. It reports an 8.4% average gain over Hy-Embodied-0.5 MoT-2B and 4.4% over Qwen3.6-A3B, and says it lands close to the previous system that activated 32B parameters. Those are publisher numbers. Treat navigation results in particular as setting-specific; the R2R-CE claim is for an RGB-only configuration.

Getting it running

  1. Use vLLM with the pinned versions Tencent validated (vllm==0.14.1, transformers==4.57.6, torch==2.9.1) and install the repository plugin, which registers the model plus its hunyuan_v3 reasoning parser and hy_v3 tool-call parser.
  2. A single node with eight 80 GB GPUs is enough for BF16.
  3. Toggle enable_thinking per request: off for fast reactive control, on for multi-step planning.

For a pilot, record a few dozen real camera frames from your robot, write the questions your planner actually asks, and compare thinking-off latency against the control loop you need. Then compare answers with a general model such as Qwen3.6 35B-A3B, which has a similar active size. Its language backbone family is Hy3; other Tencent releases are on the Tencent provider page.

How to evaluate Hy-Embodied-VLM-1.0

Compare the model on the work you need to complete.

Tasks and inputs

Hy-Embodied-VLM-1.0 is listed for Image-text-to-text, Reinforcement learning. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 32.77K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

About 3B per token, out of roughly 30B total, using 8 of 128 experts plus one shared expert.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

sensetime

SenseNova-Loop-9B

SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.

524.29K context

Tencent

Hy4 Flash

Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.

1.05M context

shanghai-ai-lab

Sage S2.1 Flash

Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.

— context

Meituan

LongCat-2.5

LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.

— context

sensetime

SenseNova-Loop-9B-Preview

SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.

262.14K context

Liquid AI

LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.

— context