Tencent has launched its new Hy3 language model and immediately released it as open-source software. The company claims its performance matches that of models with two to five times as many parameters. Anyone can download the model weights and use them commercially under the Apache 2.0 license. Tencent is also offering free API access via the OpenRouter platform for two weeks following the release.
Small but smart model
Hy3 is based on a Mixture‑of‑Experts architecture. It has 295 billion parameters, but activates only 21 billion of them for each query. It also includes a special layer with 3.8 billion parameters for so-called multi-token predictions. The model supports a context length of up to 256,000 tokens and can combine fast and slow reasoning in several modes.
Interestingly, it is a smaller model than Tencent's previous flagship, which exceeded 400 billion parameters. Even so, the company boasts better results in programming and autonomous agent tasks. According to Tencent, the range of around 300 billion parameters appears to offer the best trade-off between capabilities and cost.
Nine out of ten tasks and fewer hallucinations
Tencent's strongest argument is reliability. On the company's WorkBuddy platform, Hy3 achieves a 90 percent task completion rate. The developers focused on preventing the model from inventing non-existent facts.
Tencent based the training on a simple rule: answer only when supported by data, and admit when evidence is lacking. This has reportedly led to a significant reduction in errors, fabricated information, and logical inconsistencies. The team behind the WorkBuddy tool cites another practical advantage. When processing documents, Hy3 used 47 percent fewer tokens than the competing GLM‑5.2 model, which means lower costs for the same work.
Even catches up with GPT‑5.5 in tests
Tencent has published a set of benchmark results. Hy3 scored 78.0 in the SWE‑Bench Verified coding test and 57.9 in SWE‑Bench Pro. In the BrowseComp search test, it achieved 84.2, matching OpenAI's GPT‑5.5.
The scientific domain is attracting the most attention. In FrontierScience‑Olympiad, a test that measures scientific research capabilities, Hy3 even outperformed GPT‑5.5. In the demanding GPQA Diamond science test, it scored 90.4 points, approaching GPT‑5.5's 93 points. It remains weaker in one discipline: in Tencent's own CL‑bench test, which assesses the ability to learn entirely new material directly from context, it finished second with 23.8 points, behind Claude Opus 4.8.
Integration into tools
Hy3 has been integrated into a wide range of Tencent's own products, from the Yuanbao assistant and CodeBuddy coding tool to QQ Browser, WeChat, ima, and WeGame. Nearly 50 more applications are set to integrate it.
Thanks to Hy3, the Yuanbao assistant now supports agentic features. Users describe what they need in everyday language, and the tool creates a PowerPoint, Word, Excel, PDF, or HTML document for them free of charge. In Path of Exile: Advent on the WeGame platform, Hy3 powers an in-game assistant designed to better understand players' requests while producing fewer hallucinations.
Open source and low cost
It costs roughly 14 cents per million input tokens, about 56 cents per million output tokens, and only around four cents for cached data. That is a fraction of what closed models charge.
The model weights are available on Hugging Face and ModelScope, including an FP8 variant that roughly halves graphics card memory requirements. The model will also gradually become available on other developer platforms such as OpenRouter, Kilo, Cline, and Cherry Studio. Shortly after its release, Nous Research and Kilo Code offered free access to Hy3.
The entire project is led by Tencent's chief AI scientist, Yao Shunyu, whom the company hired late last year. From rebuilding its research infrastructure in late January through the April preview version to the current release, Tencent completed the entire development cycle in less than six months. Hy3's daily token usage has meanwhile increased twentyfold compared with the preview version.
Source: hy.tencent.com
