On August 28, Tencent's Hunyuan AI team released Hy4 preview, a 770-billion-parameter open-weight language model with a one-million-token context window. The weights are on Hugging Face, ModelScope, and CNB under an Apache 2.0 license. That means you can download all 1.56 terabytes, run it on your own servers, modify it, and never send Tencent a usage log.
Hy4 uses a mixture-of-experts architecture—770 billion total parameters, but only 49 billion active per request, which keeps inference costs closer to a 50B dense model. Tencent says it beats their previous Hy3 (released in July) and competes with proprietary models from OpenAI and Anthropic on coding, office tasks, and scientific research.
The release includes an FP8 quantized version for smaller deployments and API access via Tencent Cloud or OpenRouter for those who'd rather rent. But the weights are there, which is the point. You can audit the training, fine-tune for your use case, or run it airgapped. OpenAI's GPT-4 and Anthropic's Claude don't offer that option—you use their API or you don't use the model.
This is China's third major open-weight release in three months, following DeepSeek-V3 and Alibaba's Qwen updates. The pattern is deliberate: while US labs race to lock models behind API walls and defend training data in court, Chinese labs are shipping the artifacts. If you're deciding whether to build on rented intelligence or owned weights, the menu just expanded.