Tencent has released Hy3, an open language model from the Hunyuan family built on a Mixture-of-Experts architecture, with 295 billion total parameters and 21 billion active on each query. The weights landed on Hugging Face on July 6, 2026 under an Apache 2.0 license, making Hy3 one of the most freely licensed large models out of China and a direct rival to Zhipu/Z.ai's open GLM family.
Key takeaways
- Hy3 is an MoE model: 295B total parameters, 21B active, plus a 3.8B-parameter MTP layer
- The Apache 2.0 license grants full rights to commercial use, modification, and redistribution of the weights
- Architecture: 192 experts with 8 activated per token, 80 layers, 256K-token context, BF16 precision
- In a blind evaluation by 270 experts, Hy3 scored 2.67 out of 4, ahead of GLM-5.1 at 2.51
- The hallucination rate in internal testing dropped from 12.5 percent to 5.4 percent
What exactly Tencent shipped
Hy3 is a sparse MoE model — of a pool of 192 experts, only 8 fire on each token, so out of 295 billion parameters just 21 billion are active at any moment. On top of that sits a Multi-Token Prediction layer of 3.8B parameters that speeds up generation by predicting several tokens at once. The model has 80 layers, GQA attention with 64 heads and 8 key-value heads, a hidden size of 4096, and a 256,000-token context window. The weights ship in BF16, alongside an Hy3-FP8 variant quantized for cheaper serving.
This is the production version following April's Hy3 Preview. Tencent says it gathered feedback from more than 50 of its own products and scaled up post-training with higher-quality data and reinforcement learning. The model runs in two reasoning modes — direct, and a "deep" chain-of-thought toggled by a parameter — letting teams tune inference cost to task difficulty. To serve the full model, Tencent recommends 8 H20-3e-class GPUs, with ready recipes for vLLM and SGLang.
Why Apache 2.0 matters
The key signal here is not the parameter count but the license. Apache 2.0 is one of the most permissive open-source licenses — it allows commercial use, modification, distribution, and deployment inside closed products, with no clauses restricting user scale or region. For companies building their own deployments, that is the difference between a model you can genuinely self-host and one that ties your hands with terms of use.
This marks an interesting turn in China's open-model market. Z.ai's competing GLM (formerly Zhipu) has also embraced permissiveness — the newer GLM-5.2 shipped in June 2026 under an MIT license. Both firms are moving away from their own restrictive licenses toward standards developers know from the Western ecosystem. The contest is shifting from "open at all" to "how freely open, and how good."
Positioning against GLM and other open models
Tencent's pitch rests on the claim that Hy3 matches models two to five times its active size. The strongest evidence it offers comes from outside public benchmarks — a blind evaluation run by 270 experts on tasks from their own work. Hy3 scored 2.67 out of 4 there, ahead of GLM-5.1 at 2.51, with the biggest edge in frontend development, data-and-storage work, and CI/CD tasks.
That figure deserves a careful read. Tencent benchmarked against GLM-5.1, while Z.ai has already shipped a newer generation — GLM-5.2. The comparison therefore leaves out the freshest rival.
On public benchmarks, Hy3 lands strongest on reasoning and code work. On the hard scientific-knowledge test GPQA Diamond it reaches 90.4, and on SWE-bench Verified — which checks real code fixes — it resolves 78 percent of tasks, with variance across agent scaffoldings such as CodeBuddy, Cline, and KiloCode staying within 4 percent. Tencent itself cautions that "public benchmark scores don't tell the full story," putting its heaviest emphasis on practical usefulness and reduced hallucination.
Why it matters
The Hy3 release shows that the race among Chinese open models has stopped being a bidding war over parameter counts and become a game of usefulness and licensing freedom. Sparse MoE means a model with 295 billion parameters of capacity can be served at a cost close to a far smaller one, because only 21 billion are active. For many teams that economics outweighs a few points in a benchmark table.
The Apache 2.0 license, meanwhile, removes the deployment barrier that previously scared companies away from Chinese weights. Combined with GLM's move to MIT, it pressures the whole market — Western labs included — to compete not only on quality but on access terms. For IT teams weighing per-token costs at closed providers against their own infrastructure, each strong model on a permissive license nudges the balance toward self-hosting.
What's next
- Tencent says support for Hy3 is planned on platforms like OpenRouter and Cline — making it easier to test the model without standing up your own infrastructure
- The model is already built into Tencent's own products, including WorkBuddy, Yuanbao, and WeChat, plus an in-game assistant for "Path of Exile: Advent" — pointing to a real deployment path, not just a weight drop
- The next real test will be Hy3 against GLM-5.2 on shared benchmarks, which Tencent's official materials do not yet cover
