腾讯发布开源 Hy4 preview 模型，770B 参数、100 万 token 上下文

Tencent has released Hy4 preview, a 770-billion-parameter flagship model with 49 billion active parameters and a one-million-token context window. The company calls it its most capable model so far and says it delivers the largest generation-over-generation gain it has measured. The open-source release targets long-horizon software engineering, document-heavy office work, game development and scientific research.

We compressed Hy4-preview from 1.5TB to ～200GiB GGUF and it still works well !
Meet MIX-STQ1_0.The trick isn’t just going low, it’s deciding where: calibration data picks each layer’s bit-width, some down to 1.31-bit STQ1_0, some up to 2.06-bit IQ2_XXS. Same budget, lower… https://t.co/HItkQ1TSxA pic.twitter.com/xK8d3BH5br
— Tencent Hy (@TencentHunyuan) August 29, 2026

For software projects, Hy4 preview is designed to understand, plan, debug and verify work across extended sessions, with added focus on the visual quality of front-end output. It can also turn scattered files into documents, spreadsheets and presentations, including work involving data analysis, equations and financial models. Tencent says the model can build playable prototypes from one prompt and continue refining complex projects in game engines over multiple turns. Its research scope includes AI, condensed matter physics and pure mathematics.

In one test, Hy4 preview managed several Codex sessions in parallel and changed its research direction as results arrived. On a small-model post-training task, it coordinated work across several evaluation targets and beat Codex working independently on all eight benchmarks, according to Tencent. The company presents this as evidence that the model can judge research direction, organize experiments, and iterate on complex R&D work.

A blind internal comparison involved 163 Tencent experts rating outputs across 203 engineering tasks. Hy4 preview averaged 2.99, slightly above GLM 5.3 (2.92) and Kimi K3 (2.94). Against GLM 5.3, it recorded 46.8% wins, 12.8% ties, and 40.4% losses. Against Kimi K3, it posted 51.2% wins, 7.9% ties and 40.9% losses.

Tencent built Hy4 preview with training data shaped around work performed by its software engineers, game developers, finance analysts, and security experts. It continues to co-design the model with CodeBuddy and WorkBuddy, while access is available through Tencent Cloud, OpenRouter, Yuanbao, ima, and public repositories. API pricing per million tokens is $0.042 for cached input, $0.834 for input, and $2.501 for output. Tencent describes Hy4 preview as an early version and acknowledges that it can reason longer than needed on complex tasks and over-verify its own work.

Source