⚡ Uncle Cat AI Radar
ModelsOpen SourceAgents

Tencent releases Hy4 weights with 770B parameters

Tencent has published Hy4’s open weights and system card, giving developers access to a 770B-parameter multimodal model with a million-token context.

Tencent has released the open weights and system card for Hy4, its flagship multimodal mixture-of-experts model, marking a significant expansion of China’s open-model supply. The model contains 770 billion total parameters, activates 49 billion per token, and supports a context window of roughly one million tokens.

What changed

Hy4 accepts both text and images, supports tool calling and structured output, and is aimed at coding, computer use, document work and longer agentic tasks. Tencent’s release materials emphasize defensive cybersecurity capability, while describing safeguards intended to limit offensive exploitation assistance. The company has also made the model available through Together AI, allowing developers to test hosted inference without operating the full model themselves.

The release follows Tencent’s November 4 product launch, when the company described Hy4 as open-source but said the weights and system card would arrive later. The distinction matters: today’s event is the point at which the model becomes independently inspectable and deployable, rather than merely available through Tencent-controlled products and APIs.

Why it matters

At 770B parameters, Hy4 is not a lightweight local model. Its practical audience will initially be cloud providers, research groups and well-funded enterprises with substantial memory and inference budgets. The million-token context and native vision support nevertheless make it a credible open-weight option for long documents, software repositories and multimodal workflows.

Together AI’s same-day hosting lowers the barrier to evaluation, while the system card gives safety researchers more material to examine. The unresolved question is whether Hy4’s benchmark claims translate into dependable performance outside Tencent’s product stack, and how permissive and operationally useful the released license and artifacts prove to be.

Sources