Tencent launches Hy4 Preview, a 770B open model with a 1 M token context, outperforms Z.AI and Moonshot in internal tests
| Source: Techmeme | Original article
Tencent has launched Hy4 Preview, a 770‑billion‑parameter open model with a 1‑million token context window, claiming it outperforms Z.AI and Moonshot AI in internal tests.
Tencent has unveiled Hy4 Preview, a new foundation model that it describes as a 770‑billion‑parameter system with a one‑million‑token context window. In internal benchmarking, the company says the model outperforms competing offerings from Z.AI Co. and Moonshot AI.
The release marks a significant step for the Chinese tech giant, which has been expanding its AI portfolio beyond its well‑known cloud services. A model of this scale, coupled with an unprecedented context length, promises to handle far longer documents, codebases, or multimodal inputs without the need for chunking or external memory tricks. If the performance claims hold up in independent tests, Hy4 could narrow the gap between Tencent’s research and the leading closed‑source systems from the United States and Europe, while remaining openly accessible to developers.
The open nature of Hy4 is likely to accelerate experimentation across the Nordic AI community, where researchers and startups frequently adopt publicly available models to build niche applications. A longer context window opens possibilities for advanced summarisation, legal‑document analysis, and complex reasoning tasks that have previously strained existing models.
What to watch next is whether third‑party evaluations confirm Tencent’s internal results and how quickly the model is integrated into popular frameworks and APIs. Attention will also turn to the licensing terms Tencent attaches to Hy4, the speed of ecosystem tooling, and any strategic responses from Z.AI, Moonshot and other rivals. The broader market will gauge whether the model’s capabilities translate into real‑world performance gains, potentially reshaping the competitive landscape of large‑scale open AI models.
Sources
Back to AIPULSEN