Tencent officially released Hy3 on July 6, 2026, an open-source Mixture-of-Experts language model with 295 billion total parameters and 21 billion active per token, supporting a 256K-token context window. Unlike the model's April preview, which excluded the EU, UK and South Korea from its license, the official release carries an unrestricted Apache 2.0 license, giving developers worldwide freedom to self-host, fine-tune and build commercial products with it.
Tencent's research team reported strong benchmark results: 78.0 on SWE-Bench Verified for coding and 90.4 on GPQA Diamond for reasoning, while hallucination rates fell from 12.5% to 5.4% and multi-turn dialogue accuracy rose from 42.9% to 75.1%. In blind expert testing the model outscored rival GLM-5.1, and reviewers noted it matches larger flagship models despite its smaller active-parameter footprint.
Hy3 already powers Tencent products including WorkBuddy, Yuanbao, Marvis and ima, and is rolling out to third-party platforms such as Hugging Face and OpenRouter, where it is free to use through July 21, 2026. The release adds to a wave of competitive open Chinese models — which have accounted for as much as 46% of weekly AI token usage among US companies in recent months.