Tencent HY-Embodied-0.5

HY-Embodied-0.5 is the first model in the family that Tencent Robotics X builds its robot software on. It is not a controller: it looks at what a camera sees, reasons about the space and the objects in it, and decides what should happen next. The motor commands come from a separate model that sits on top of it. The published weights are the MoT-2B variant, about 3.8 billion parameters in the file. Tencent describes it as a four-billion-parameter model of which only 2.2 billion are computed for any given token, using an architecture the company calls Mixture-of-Transformers: the vision pathway gets its own dedicated computation instead of sharing one stack with the text pathway. The stated result is the speed of a small dense model with the perception of a larger one, which is the trade every robot builder wants, because a machine that pauses to think has already dropped what it was holding. A second variant, 32B, exists but its weights were not published. Tencent compares it to Gemini 3.0 Pro and says the reasoning of the large model was distilled down into the small one that was released. The claim about the 32B model cannot be checked outside the company. The training set is the interesting part of the recipe: over 100 million data points specific to physical space and embodied action, and more than 200 billion tokens. On the vendor's own table of 22 benchmarks the 2B model leads models of similar size on most of them, including spatial ones — 66.3 on MindCube against 28.4 for the nearest comparable Qwen model. These are the producer's figures; no independent evaluation was found. The licence matters more than the benchmarks for a European reader. This model does not ship under Apache 2.0 but under the Tencent HY Community License, whose opening line states in capitals that the agreement does not apply in the European Union, the United Kingdom and South Korea. A company in those territories is not offered restrictive terms — it is offered none. The two action models built on this same backbone are, by contrast, plain Apache 2.0. The technical report is arXiv 2604.07430. The repository went up on 2 April 2026 and the release was announced on 9 April.

▸Videos

No videos yet.

▸Reviews