Tencent Hunyuan Releases Hy4 Preview with 770B Parameters and 1M Context

Tencent Hunyuan has released Hy4 preview, an open large language model aimed at productivity-oriented use cases. According to the release information, the model has 770B total parameters, 4.9B activated parameters, and a context length of 1M. Hunyuan says the model is positioned among the leading open models for coding, office, and scientific tasks.

Key points

• Large total capacity with limited activation: Hy4 preview separates the model’s total parameter capacity from the parameters activated during an inference pass. Its 770B total parameters and 4.9B activated parameters suggest an architecture designed to retain broad capacity while limiting the active computation for each request. The figures alone, however, do not establish deployment cost or performance.
• A 1M context window: The stated context length is intended to support long codebases, large documents, and extended task histories. In practice, the usefulness of a long context depends on serving infrastructure, information selection, and the model’s ability to preserve and retrieve relevant details across a very long input.
• Multiple distribution channels: Hy4 preview has been made available through Hugging Face, GitHub, ModelScope, and Gitcode. It is also listed on Tencent Cloud TokenHub and OpenRouter, giving developers options ranging from direct experimentation to hosted API integration.
• An agent-oriented direction: The release puts notable emphasis on applications beyond text generation. Demonstrations involving game engines and MCP-related tooling point toward a model that can participate in software workflows, interact with tools, and help move a task forward inside an existing environment.

Why it matters

Hy4 preview is significant not only because of its headline parameter count. It combines long-context support, open distribution, and a product narrative centered on agents. For developers, connections to game engines and MCP-compatible toolchains could make it easier to place the model inside applications that already have structured actions, resources, and execution environments.

The available material remains limited, however. The claim that Hy4 preview belongs to the top tier of open models cannot be independently assessed from the announcement summary alone. More information is still needed on evaluation methodology, licensing, hardware requirements, serving costs, and the differences between local and hosted access. Developers should therefore validate the model against their own workloads rather than relying only on scale or context-window figures.

The broader signal is clear: competition among large models is moving beyond parameter counts. Open ecosystems, long-context reliability, tool use, and practical agent deployment are becoming equally important parts of a model release.

Source: OSChina