Tencent opens a 770B coding model

Tencent released and open-sourced Hy4 preview on August 28, a mixture-of-experts model with 770 billion total parameters, 49 billion active per token, and a context window past 1 million tokens. It roughly doubles the parameter count of its predecessor, Hy3, and is aimed squarely at coding, office work, data analysis, and scientific research rather than general chat.

Tencent’s own blind evaluation, 163 human experts scoring 203 engineering tasks, put Hy4 preview at 2.99 out of 4.00, edging out GLM-5.3 at 2.92 and Kimi K3 at 2.94. Self-reported benchmarks from the model’s own maker deserve the usual skepticism, but the margin over two of the strongest open competitors of the moment is narrow enough to be plausible rather than promotional.

What you can actually use today

The model is live now inside Tencent’s WorkBuddy and CodeBuddy apps, plus Yuanbao and ima, with API access through Tencent Cloud TokenHub and OpenRouter. API pricing lands under a dollar per million tokens in, about 83 cents, and around $2.50 per million tokens out, with cache hits priced far lower still, roughly 4 cents per million. WorkBuddy and CodeBuddy are free for the first two weeks after launch, and Hy3 access stays free through September 30 for anyone still on the older model.

Tencent also says Hy4 preview was the first of its models to help optimize its own training methods, and that the resulting inference-system tuning delivered a 31.8% throughput gain. That claim is unverified outside Tencent’s own report, but it fits a pattern this year of labs using their newest model to speed up the infrastructure that serves it, Anthropic’s automated alignment researcher being the clearest recent example on the safety side rather than the infrastructure side.

For operators evaluating open-weight options, the practical test is simple: run your own coding and office-automation workload against Hy4 preview’s free two-week window before the next contract renewal, and weigh the sub-$1 input price against whatever you’re paying a closed-model vendor for comparable work.