Tencent Releases Smaller Open Model Hy3, Matches Larger Rivals in Scientific Tasks
Tencent's Hy3 demonstrates that a smaller open model can compete with much larger competitors in scientific reasoning, challenging the prevailing trend of scaling model size.
Reporting from 1 source: GIGAZINE.
Tencent released Hy3, an open-source AI model with 295 billion total parameters and 21 billion active parameters, on July 6. The model matches or outperforms larger models like GLM-5.2 and DeepSeek-V4-Pro in benchmarks, and beats GPT-5.5 in scientific research tasks. Hy3 is designed to reduce hallucinations by only answering when evidence exists. It is available under Apache 2.0, with API pricing from 1 yuan per million input tokens, and free on OpenRouter for two weeks.
Tencent's Hy3 follows a different design philosophy from many current large models. The developers implemented data cleaning and training constraints that instruct the model to only answer when it has evidence, and to state the lack of evidence when it does not. According to the company, this reduces confusion of facts, fabrication, and logical contradictions. The result is a 295-billion-parameter model that, in its internal document processing tests, used 47.4% fewer tokens than the larger GLM-5.2 model.
Hy3 is released as an open model under the Apache License 2.0, and an API is available at 1 yuan per million input tokens. For a limited two-week period, access is free on OpenRouter.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.