Tencent open-sources the 770B-parameter Hy4 preview

Background

Tencent's Hy3 model had been launched on 6 July 2026. The Hy4 preview came less than two months later, which Caixin described as a faster pace of major upgrades.^[1] A preview is an early version made available for testing before a final release.

What happened

On 28 August 2026 Tencent released and open-sourced Hy4 preview, a model tailored for productivity tasks.^[1] Because it is labelled a preview, it is an early release of the Hy4 model rather than a final version. The model has 770 billion total parameters and 49 billion active parameters, and its context exceeds one million tokens.^[1]

Weights, including an FP8 version, were released under the Apache License 2.0.^[2] Apache 2.0 allows use, modification and redistribution of the released weights.

Key details

Tencent reported 65.7 on SWE-bench Pro and 85.4 on Terminal-Bench 2.1.^[2] SWE-bench Pro tests models on software engineering problems, and Terminal-Bench measures how well a model completes tasks in a command-line environment.

Caixin placed Hy4 against rivals including Moonshot's Kimi K3, DeepSeek V4 Pro and Z.ai's GLM-5.3.^[1] Comparisons of this kind show which models the release was measured against in press coverage.

Why it matters

The preview is larger than Hy3, which had 295 billion total parameters.^[1][2] Because the weights and FP8 version were published under Apache 2.0, the model can be examined and built on by anyone who downloads it under those terms.^[2] The release is also a reference point for readers comparing open-weight models, since the license, the one-million-token context and the parameter counts are all stated in it.^[1][2] A preview is an early version made available for testing, and details can change before a final release. Tencent's description of the model as tailored for productivity tasks points to work-oriented use, and the benchmarks in the release, such as SWE-bench Pro and Terminal-Bench 2.1, test software engineering and command-line work.^[1][2]