Tencent open-sources 770B-parameter Hy4 preview
Tencent has open-sourced Hy4 preview, a 770B-parameter mixture-of-experts model with a context window over 1M tokens and low API pricing, though its only performance evidence comes from Tencent's own internal evaluation.
Tencent announced the release of Hy4 preview, a large language model with 770B total parameters, 49B of them active, and a context window of more than 1M tokens. The company is targeting productivity work: coding, office tasks, game development and scientific research.
The weights are open-source. The model is also available in Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima, and through APIs on Tencent Cloud TokenHub and OpenRouter. API pricing is $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached tokens. Use on WorkBuddy and CodeBuddy is free for two weeks after launch.
Tencent's only performance figures come from an internal blind evaluation with 163 experts and 203 engineering tasks. Hy4 preview averaged 2.99 out of 4.00, a little ahead of Kimi K3 at 2.94 and GLM-5.3 at 2.92. Tencent did not publish results on public benchmarks.
The company also says Hy4 preview helped build itself. According to Tencent, the model proposed and ran experiments on training methods, data strategies, evaluation frameworks and low-level operators, which the company calls "an early-stage recursive self-improvement loop." Tencent also says the model optimized its own inference stack, including operator fusion and communication, and raised end-to-end throughput by 31.8% over a baseline. These claims have not been independently verified, and Tencent did not say what the baseline was.
Tencent says more models in the Hy4 series are coming soon. It has not given a date for the official, non-preview release.