[DigitalToday reporter Chi-gyu Hwang (황치규)] Tencent has released the official version of its in-house large language model Hy3 under an Apache 2.0 license, VentureBeat reported on Sunday.

Hy3 is a 295 billion-parameter mixture-of-experts model with 21 billion active parameters, the report said. A preview version released in April excluded the European Union, Britain and South Korea from its license. The official version no longer has that restriction.

The open-source community also focused on the license change itself. Tencent will provide Hy3 for free for 2 weeks via OpenRouter.

Tencent said it fixed task execution and interaction issues after receiving feedback from more than 50 product teams since the late-April preview.

The structure remains unchanged. It activates the top 8 out of 192 experts, and its multi-token prediction layer has 3.8 billion parameters. Its context window is 256,000 tokens.

Tencent said Hy3 scored higher than GLM-5.1 in a blind test involving 270 experts. On a 4-point scale, Hy3 scored 2.67 and GLM-5.1 scored 2.51. GLM-5.2, the latest model released by Z.ai in June, outperformed Hy3 on most coding benchmarks. It scored 84.2 versus 78.0 on SWE-bench Verified and 81 versus 71.7 on Terminal-Bench 2.1.

Hy3 showed strengths in search and tool use. It scored 84.2 on BrowseComp and 91.0 on DeepSearchQA, the best among the open models Tencent compared. VentureBeat said it was competitive with Claude Opus 4.8 and GPT-5.5.