Tencent has released a very large open model, but raw parameter count is the least interesting part of the announcement. Hy4 Preview contains 770 billion total parameters, activates 49 billion for each token, and supports more than one million tokens of context. That is a sparse system designed to bring enormous capacity into a more controlled inference budget.

The model is available through Tencent products including WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access through Tencent Cloud TokenHub and OpenRouter. Tencent also published the model as open source. Preview matters here. This is an operating release and an invitation to test, not a finished claim that every edge case has been solved.

Tencent says an internal blind evaluation used 163 experts and 203 engineering tasks. The company reports a 2.99 score out of four for Hy4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Useful signal, limited evidence. Internal evaluations can reveal where a builder has confidence, but they do not replace independent replication across workloads the company did not choose.

The stronger systems claim is recursive. Tencent says Hy4 participated in optimizing its own training methods, data construction, evaluations, and low-level operations. It also reports that operator fusion and communications work increased end-to-end throughput by 31.8 percent against its baseline. Again, those figures come from Tencent, but they point toward a more important product category: models improving the machinery used to build and serve models.

Pricing starts at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached tokens. The market will decide whether that combination of scale, sparsity, openness, and cost survives real production traffic. The useful question is not whether 770 billion sounds impressive. It is whether Hy4 can make the surrounding engineering loop measurably faster without requiring a heroic operations team to keep it alive.

LaunchPad positionThe scale is notable, but the strategic signal is Tencent's claim that the model helped optimize training methods, data, evaluations, and low-level operations. Those results remain company-reported and the model is explicitly a preview.
Reporting standard

This report draws on the linked primary sources and reputable reporting. Company statements are treated as claims until independently demonstrated.