Tencent Releases Hy4 Preview with 770B Parameters

Tencent has launched Hy4 Preview, a new open-weight text-only large language model. The model represents a massive scale-up from its predecessor, Hy3, which was released in July. Hy4 Preview features 770 billion total parameters and 49 billion active parameters, compared to Hy3's 295 billion total and 21 billion active parameters. The new model also expands its context window to 1 million tokens, up from the 256,000 tokens supported by Hy3. The model's weights are available on Hugging Face, where the download size is a massive 1.56 terabytes, a significant increase from the 598 gigabytes required for Hy3.

An analysis of the model's chat template on Hugging Face reveals a new reasoning control mechanism. Developers can configure the reasoning effort of the model using a specific parameter that accepts two distinct settings. The default setting is high, which enables the model's full reasoning capabilities. Alternatively, developers can set the parameter to no_think, which disables the reasoning process entirely. This binary choice gives practitioners direct control over whether the model spends compute on internal chain-of-thought processing before generating a final response.

For practitioners, this dual-mode setup offers a way to balance latency and output quality. When high reasoning is enabled, the model generates a hidden reasoning trace. Early tests of this trace show that the model uses slightly truncated English during its internal thinking process, likely to optimize token efficiency. For example, when prompted to generate SVG code, the model's internal monologue weighed design choices in short, fragmented sentences. By choosing between high and no_think, developers can tailor the model's behavior to suit either complex problem-solving tasks or faster, more straightforward text generation needs.