Tencent has announced the release of Hy4 preview as an open model. This is a MoE model with 770 billion total parameters and 49 billion active parameters. It supports a maximum context length of 1 million tokens.
Hy4 preview reportedly outperformed GPT-5.6 Sol in SWE Atlas Refactoring, which measures agent performance. It also surpassed the same model in SWE-bench Pro. Tencent stated that the model features improved capabilities in long-term task understanding, planning, and verification, making it capable of handling a wide range of tasks such as documentation and game development.
The model is released under the Apache License 2.0 and is available for free. Running the original model requires approximately 1.5TB of VRAM.
Tencent has also released quantized build that minimizes performance degradation. The STQ1_0 version, which employs a method of appropriately selecting quantization targets, can reduce VRAM usage to approximately 214GiB. Tencent stated that they have kept the performance gap small relative to the model size.
Source: Tencent releases AI model "Hy4 preview," outperforming GPT-5.6 Sol in some tests & releases 1-bit quantized build lightweight version down to 213GB (GIGAZINE, 2026-08-31)