Advertisements

Tencent has released downloadable weights for Hy4, its multimodal flagship model, alongside deployment instructions and a system-card link. The November 20 announcement gives developers a way to run the model themselves and inspect its documented architecture. The release includes an instruction-tuned model, an FP8 version and pretrained base weights. [1][2]

Hy4 accepts text, images and video. Tencent lists a language backbone of 770 billion parameters, with 49 billion activated per token, and a context limit of 1,048,576 tokens. Those figures describe different parts of the system: active parameters concern the computation selected for a token, while the complete model still has to be stored and served. [1]

Advertisements

What the release contains

The backbone uses a mixture-of-experts architecture. Tencent also supplies a vision encoder and a DFly2 component for speculative decoding, which drafts tokens to speed generation. Its published figures exclude those components from the 770-billion-parameter backbone total. Readers comparing headline model sizes should account for that distinction. [1]

The base checkpoint serves a different role from the chat model. Tencent's model card identifies it as pretrained rather than instruction-tuned, and directs chat and reasoning examples to Hy4 and Hy4-FP8. It also links to a GGUF compression release for testing and provides a fine-tuning guide. [2]

The hardware requirement is substantial

The vLLM recipe starts with an eight-GPU NVIDIA node. It estimates roughly 944 GB of aggregate GPU memory for FP8 and around 1,859 GB for BF16. The listed checkpoint sizes on disk are about 786 GB and 1,549 GB respectively. These are server-scale requirements; the number of active parameters alone is a poor guide to whether the download will fit a workstation. [3]

Advertisements

Context capacity also needs a closer look. Although the checkpoint advertises a one-million-token window, the vLLM recipe begins at 131,072 tokens and says a full one-million-token request has not been validated for that recipe. Its initial multimodal setting permits one image and disables video. Operators need to distinguish model capabilities from the settings actually enabled in a serving configuration. [3]

A custom license and documented limits

Hy4 uses the Tencent Hy License, Version 1.0. The license says it derives from Apache 2.0 with modifications, including additional commercial terms for large model-as-a-service businesses. It is a separate license, so the Apache name should not be treated as a shortcut for understanding the release’s conditions. [4]

Section 10 requires a separate written license for qualifying MaaS operators above specified revenue or user thresholds. It describes a US$50 million worldwide gross-revenue threshold over a consecutive 12-month period, or a relevant product or service exceeding 100 million monthly active users. The section also defines MaaS and an internal-use exception. [4]

Tencent’s own limitations section says Hy4 still trails leading proprietary models on difficult research mathematics, long coding tasks, model self-improvement and advanced visual question answering. Its published benchmark results are vendor evaluations. For a deployment decision, the useful next step is a trial on representative workloads, with the serving configuration and license reviewed together. [2]

Sources

[1] Tencent Hy4 official repository

[2] Tencent Hy4 model card

[3] vLLM Hy4 serving recipe

[4] Tencent Hy License, Version 1.0

Illustrative photograph of code on a screen; not a Hy4 interface or benchmark. Slashme / Wikimedia Commons. CC0 1.0. Wikimedia 1280-pixel rendition. Image source. Image license / permission.

Advertisements

Have a correction or a useful perspective?
Get in touch with the publication.