Tencent open-sources Hy4 preview with 770B parameters and a million-token context window

Tencent has released and open-sourced Hy4 preview, positioning the new Hunyuan model as a major step in China’s fast-moving open-weight AI race. In an official announcement dated August 28, 2026, the company said Hy4 preview has 770 billion total parameters, activates 49 billion parameters per token and supports a context window exceeding one million tokens.

The release is aimed squarely at productivity work rather than a narrow benchmark demonstration. Tencent said the model is designed for coding, office work, analysis, game development and scientific research. Hy4 preview is available as an open-source model and is also being rolled out through Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima. Developers can access it through Tencent Cloud TokenHub and OpenRouter, while users of WorkBuddy and CodeBuddy receive a two-week free trial period. Tencent also extended free use of Hy3 on those products until September 30.

The specifications show how quickly the open model market is scaling. A 770 billion-parameter mixture-of-experts model is not intended to run casually on consumer hardware, but the active-parameter count allows each token to use a smaller portion of the network than the headline size suggests. The million-token context window is especially important for long software projects, large document collections and multi-file analytical tasks, where the ability to keep more material in context can matter as much as raw reasoning capability.

Tencent said Hy4 preview was expanded across model size, context length and training data, with gains from both pre-training and post-training. The company also said it built training data with internal domain experts, including software engineers, game developers, finance analysts and security specialists. That approach reflects a broader industry move away from generic chat benchmarks toward models tuned on the messy workflows that professionals actually run: debugging, reviewing, preparing reports, coordinating files and converting analysis into documents, spreadsheets or presentations.

The company highlighted internal blind evaluations involving 163 experts and 203 engineering tasks. According to Tencent, Hy4 preview scored an average of 2.99 out of 4.00, slightly ahead of GLM-5.3 and Kimi K3 in that test. Because the evaluation is internal, outside developers will still need to test the model on their own workloads before drawing firm conclusions. But the disclosed numbers are notable because they show Tencent measuring the model against other leading Chinese open-weight systems in practical engineering scenarios.

One of the more unusual claims is that Hy4 preview contributed to its own development. Tencent said the model participated in optimizing training methods, data strategies, evaluation frameworks and low-level operators, then used code, logs and feedback from experiments in subsequent rounds. The company also said the model analyzed inference bottlenecks and helped improve operator fusion and communication optimization, increasing end-to-end throughput by 31.8 percent compared with a baseline.

Pricing is another part of the strategy. Tencent listed API pricing of 0.834 dollars per million input tokens, 2.501 dollars per million output tokens and 0.042 dollars per million cache-hit tokens. That places Hy4 preview in the broader contest to make advanced reasoning and long-context systems affordable enough for routine work. The company said more Hy4-series models are expected soon, suggesting this preview is both a product release and a feedback-gathering step. For developers, the release adds another large open model to evaluate; for the industry, it signals that China’s open-weight frontier remains aggressive in scale, context length and deployment pricing.