Tencent's open-weight Hy3 model holds No. 2 on Hugging Face days after release

Tencent Hunyuan’s Hy3, an open-weight Mixture-of-Experts model released July 6, 2026, is holding the No. 2 spot on Hugging Face’s trending models list several consecutive days later, as of July 8.

Sustained placement, rather than a first-day spike, is the update here: the model has not slipped since its release earlier in the week. Hy3 is a 295-billion-parameter Mixture-of-Experts model that activates 21 billion parameters per token across 192 experts with top-8 routing, adds a 3.8-billion-parameter multi-token-prediction layer and a 256,000-token context window, and ships under the permissive Apache 2.0 license, according to the model card.

Coding performance is the draw. Tencent reports a score of 78.0 on SWE-bench Verified, a benchmark of real-world software-engineering tasks, plus 57.9 on SWE-bench Pro. The company also says it cut hallucination rates to 5.4% from 12.5% and multi-turn issue rates to 7.9% from 17.4% against its prior internal baseline, per its GitHub repository.

Those numbers are Tencent’s own and have not been independently reproduced; the internal-baseline comparison uses an unnamed prior model, so the improvement is hard to place against rivals. Weights are distributed on Hugging Face, ModelScope, GitCode and CNB, with vLLM and SGLang support for 8-GPU tensor parallelism.

Whether Hy3 stays near the top of the chart or fades in the weeks ahead is the signal to watch.