Tencent HY-Embodied-0.5-X

HY-Embodied-0.5-X is the compact end of Tencent's robotics line: a vision-language model small enough to run on the robot rather than in a data centre, published on 24 April 2026 by Tencent Robotics X and the HY Vision Team. It is not a fresh design but a reworking of HY-Embodied-0.5, released two weeks earlier, retuned specifically for physical interaction. The published weight file holds 3.79 billion parameters; Tencent describes the architecture as MoT-2B, with roughly 2 billion active at a time. The stated hardware floor is a single GPU with 16 GB of memory, which is the whole argument for this size — a machine that carries its own perception model does not need a network link to decide whether it has picked something up. Tencent claims state-of-the-art results on ten embodied task-planning benchmarks and first place among edge-side domain models on seven of them. Those are vendor figures with no independent check. Where the model is aimed is clearer than how well it scores: fine-grained manipulation understanding, spatial reasoning, action prediction, risk assessment, reference grounding, and long-horizon planning. The context window is 262,144 tokens, eight times what Tencent's later and much larger Hy-Embodied-VLM-1.0 offers — an unusual inversion, and a reminder that context length and capability are separate axes. The licence needs care. This model ships under the Tencent HY Community License Agreement, which opens with a statement in capitals that it does not apply in the European Union, the United Kingdom or South Korea. A European user is not offered restrictive terms; they are offered none. The newer Hy-Embodied-VLM-1.0 from the same team is plain Apache 2.0 with no such exclusion, so within one family Tencent runs both regimes. The licence document itself is dated 22 April 2026, two days before the public release note.

▸Videos

No videos yet.

▸Reviews