Tencent Hy3 open-source MoE model: 295B params, 21B active

Tencent has released Hy3 on Jul 6, 2026. The model uses a Mixture-of-Experts architecture with 295 billion total parameters, 21 billion active at any given time, plus 3.8 billion parameters in an added MTP layer, and it supports context lengths up to 256,000 tokens.

What is Hy3 and how is it built?

Hy3 is a Mixture-of-Experts model with 295 billion total parameters, of which 21 billion are active at any moment and 3.8 billion belong to an added MTP layer; it also accepts context windows up to 256,000 tokens. The architecture is released under an Apache 2.0 license and Tencent offers an FP8-quantized version to reduce storage and inference cost.

The Mixture-of-Experts design means only a subset of the model's parameters are used for each forward pass, which is why Tencent reports 21 billion active parameters despite the 295 billion total parameter count. Tencent lists distribution points for Hy3 on Hugging Face, ModelScope and GitHub.

How did Hy3 perform in evaluations?

In a blind evaluation conducted with 270 experts, Hy3 scored 2.67 out of 4, beating GLM-5.1 which scored 2.51; Tencent also says Hy3 "matches the performance of models two to five times its size." Internal testing cited by Tencent showed the hallucination rate dropped from 12.5 percent to 5.4 percent.

Those figures are the concrete performance data provided by Tencent. The blind-expert score is the clearest comparative number published: 2.67 for Hy3 versus 2.51 for GLM-5.1. Tencent pairs that result with the claim that Hy3 can reach parity with models two to five times larger when measured on their evaluation set.

Where is Hy3 available and how will Tencent use it?

Hy3 is available under Apache 2.0 on Hugging Face, ModelScope and GitHub, and an FP8-quantized build is also published; Tencent plans support for platforms like OpenRouter and Cline and has already integrated Hy3 into its products. The company has embedded the model in WorkBuddy, Yuanbao, WeChat and the game assistant for "Path of Exile: Advent."

Making the model available under Apache 2.0 opens it for a broad set of downstream uses without the stricter controls of some other licenses. Tencent’s note that platform support is planned and that Hy3 already powers internal products signals a dual strategy: public release for external developers and immediate internal deployment.

Why it matters

Hy3 packages a very large parameter count with a relatively small active footprint: 295 billion total parameters but 21 billion active at inference. That combination, plus the published blind-evaluation score and the reduction in hallucination rate from 12.5 percent to 5.4 percent, suggests Tencent is prioritizing cost-effective inference and improved factuality while retaining high capability. For developers and product teams seeking open models, Hy3’s Apache 2.0 licensing and FP8 quantization lower barriers to experimentation and deployment.

Enterprises building chat assistants, productivity tools or game agents may be particularly interested because Tencent has already shown internal product integrations, indicating production readiness rather than a research-only release.

What to watch

Look for independent benchmarks and third-party attestations of the 2.67 blind-expert score and the claimed parity with models two to five times larger, and for the announced platform integrations such as OpenRouter and Cline to appear. Another clear signal will be community uptake on the listed repositories and the availability of the FP8-quantized build in common inference stacks.