Chinese technology company Tencent today presented a preview of its Hy4 large language model, which accepts text input only (no vision). According to the published details, Hy4 contains 770 billion total parameters, of which 49 billion are described as active parameters. The model’s context window is specified as 1,000,000 tokens.
How it compares to Hy3
Hy4 represents a substantial increase in scale compared with Tencent’s Hy3 model released in July. Hy3 had 295 billion total parameters, 21 billion active parameters, and a context window of 256,000 tokens. Hy4 expands both total and active parameter counts and increases the context capacity to 1 million tokens.
Chat template and the "reasoning_effort" setting
A chat_template.jinja file available on Hugging Face includes logic for a reasoning_effort parameter. The template sets reasoning_effort to "high" by default if it is not defined, and only accepts two valid values: "high" and "no_think". If an invalid value is supplied, the template raises an exception.
Community testing and observed behavior
In a user test, the prompt "Generate an SVG of a pelican riding a bicycle" was executed with the default "high" reasoning mode via OpenRouter. The model produced a visible reasoning trace that considered details such as whether to add a helmet or sunglasses to the pelican and whether a helmet might obscure the bird’s long beak.
Why this matters
The larger parameter counts and the 1 million token context window increase the model’s potential for handling longer contexts and more complex tasks. The chat template’s two reasoning modes and the appearance of detailed internal reasoning in public tests indicate that Hy4’s behavior can be configured and that it can perform multi-step, self-reflective reasoning in its default mode.
Summary
Tencent’s Hy4 preview marks a technical scale-up for the company’s LLM work with 770 billion parameters, 49 billion active parameters, and a 1,000,000-token context window. Publicly available template code and community examples show two reasoning modes and suggest the model can produce explicit multi-step reasoning traces when run in the default "high" mode.
