Tencent released Hy3 on July 6, 2026—its most capable open-source AI model to date and one the company says can outperform GPT-5.5 on scientific benchmarks. Built on a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active parameters, Hy3 is available under the permissive Apache 2.0 licence and accessible for free through OpenRouter until July 21.
What Happened
Tencent’s Hunyuan team published the full release of Hy3 after an April preview that was refined based on feedback from 50 internal product teams. The final model shows meaningful improvements over that preview: hallucination rates dropped from 12.5 percent to 5.4 percent, and commonsense reasoning errors fell from 25.4 percent to 12.7 percent. The model supports a context window of up to 256,000 tokens, making it well-suited for long-document analysis, agentic workflows, and extended multi-turn conversations.
Hy3 is available immediately on Hugging Face and ModelScope, with a progressive rollout to developer platforms including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio. Tencent has also integrated the model directly into WeChat and several of its core business products, signalling that Hy3 is intended as a production deployment rather than a research preview.
Why It Matters
Hy3 is noteworthy for several reasons beyond its benchmark scores. The MoE design means that despite 295 billion total parameters, only 21 billion are active during any single inference pass. This dramatically reduces compute costs, making the model economically viable even for smaller organisations. The Apache 2.0 licence allows unrestricted commercial use without royalties—a sharp contrast to many frontier models that require enterprise licensing agreements or impose usage restrictions.
The release also extends China’s growing lead in open-source AI. Hy3 joins a strong field of Chinese open-weight models, including Meituan’s LongCat-2.0, a 1.6-trillion-parameter model built entirely on domestic Chinese chips, as evidence that Chinese technology companies are setting their own frontier rather than following Western labs. At the same time, Hy3 enters a fiercely competitive landscape where OpenAI’s GPT-5.6 Sol, Terra, and Luna currently represent the ceiling for proprietary AI performance.
Background and Context
The open-source AI movement has accelerated dramatically in 2026. Meta’s LLaMA series established the template: release powerful model weights freely, gain ecosystem adoption, and leverage that adoption to drive tooling, research, and ultimately commercial advantage. Tencent appears to be following a similar strategy with Hy3—using free access to seed global developer adoption while monetising through integrations inside its own products and cloud services.
There are legitimate questions, however, about independent verification of Hy3’s benchmark claims. The comparisons cited—including performance against GPT-5.5 and GLM-5.2—were conducted by Tencent itself, and third-party evaluations will take time to emerge. VentureBeat’s analysis noted that while Hy3 outperforms GLM-5.2 across most evaluated categories, it trails in coding tasks, which may limit its appeal for software development use cases where code generation is a priority.
“Hy3 brings the performance of a frontier model to developers who need it most—without licensing costs or access restrictions.”
Tencent Hunyuan Team, July 2026
What Comes Next
The competitive response from Western AI labs will be closely watched. An open-source model that credibly challenges GPT-5.5 at zero licensing cost puts real pressure on every organisation currently paying for proprietary API access. Enterprises evaluating AI vendors will need to weigh the flexibility and cost savings of Hy3 against the support, reliability, and integration ecosystems that closed providers offer.
For geopolitical observers, Hy3 also carries strategic significance. Chinese AI exports through open-source channels face far less regulatory scrutiny than commercial proprietary products, meaning a model of this capability distributed freely worldwide will inevitably reach users in regions where Chinese technology is otherwise restricted. How export control authorities respond—if at all—will set an important precedent for governing open-source AI globally. The rapid competitive pressure from Chinese open-source models adds further complexity for established AI labs: even market leaders like Anthropic, which recently surpassed OpenAI in annual revenue, must now contend with capable open-weight alternatives that cost nothing to access.
