Tencent Hy3
Tencent · China · 2026
Tencent's first globally released open-weights flagship — a 295-billion-parameter MoE with 21 billion active per token, three selectable thinking depths and an Apache 2.0 licence.
Hy3 is the third generation of Tencent's Hunyuan line and the first that the company pushed onto the global market under its own brand. It is a sparse mixture-of-experts model: 295 billion parameters in total, of which roughly 21 billion are activated per token, spread across 192 experts with top-8 routing over 80 layers, plus a separate 3.8-billion-parameter multi-token-prediction stack used to speed up decoding. The context window is 256,000 tokens. Unusually for a model of this size, Hy3 exposes the amount of deliberation as a user setting: no_think for immediate answers, think_low for short chains of reasoning and think_high for extended deliberation, so the same endpoint can serve both a chat reply and a long analytical task. A preview build appeared in April 2026 under a licence that excluded the European Union, the United Kingdom and South Korea. The official release on 6 July 2026 dropped those carve-outs and moved the weights to a plain Apache 2.0 licence with no regional restrictions — one of the most permissive terms attached to a model of this scale. On 5 August 2026 Tencent announced global availability across its own products and cloud: the WorkBuddy assistant, the Miora design tool and Tencent Cloud TokenHub, alongside third-party access through OpenRouter and open-weight downloads on Hugging Face and ModelScope. The company says API traffic to Hy3 ran 68 times higher than to the previous generation within a week of launch, putting it at the top of OpenRouter's usage ranking. Tencent's own comparison places Hy3 level with flagship models two to five times its size on reasoning, instruction following and agentic work; independent coverage broadly agrees, with coding the one area where rival open models still lead.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!