
Tencent released downloadable weights for its flagship Hy4 on November 20, following its earlier API launch. The release includes instruction-tuned, FP8 and pretrained base checkpoints—a new deployment option, not a new model generation. The official model card links the available versions.
Flagship Hy4 becomes a self-hosting option
Hy4 supports text, image and video understanding with a one-million-token context window. Tencent positions it for coding, visual workflows and computer use. These are vendor-described capabilities, not results from our testing. The official repository documents the model and serving options.
The pretrained base checkpoint is not an instruction-tuned chat model. Teams should distinguish it from Hy4's instruction-tuned and FP8 versions when choosing what to evaluate. This release concerns flagship Hy4, not the separate Hy4 Flash model.
Deployment requirements matter
This is not a lightweight desktop model. The current vLLM FP8 serving recipe targets an eight-GPU NVIDIA node and requires a Hy4-specific image or nightly build. That is a particular supported configuration—not evidence that any eight-GPU machine will work.
The recipe's evaluated configuration uses 131,072 context tokens; a full one-million-token request remains unvalidated for that recipe. The model's advertised context capacity and the serving configuration tested by the recipe are different things.
The weights use the custom Tencent Hy License, not Apache 2.0. Download availability should not be mistaken for identical licensing terms across open-weight models.
Why it matters for builders
Downloadable weights let teams evaluate Hy4 inside infrastructure they control, rather than relying solely on a provider's API. Potential uses include screenshot-based interface checks, supervised coding and document workflows.
For small teams such as SunMarc App Labs, the practical next step is a bounded evaluation—not a commitment to self-host. A useful trial would compare task completion, human review effort and the cost of operating the required infrastructure. We have not deployed or tested these checkpoints.
For the earlier API launch and the capabilities announced before the weights became available, see our original Hy4 coverage.