Tencent has officially launched its Hunyuan Hy3 large-scale model. In comparison to the preview iteration, this new version showcases a substantial leap in intelligence, boasting performance on par with leading models that possess 2 to 5 times more parameters. Concurrently, Hy3 introduces a more budget-friendly pricing model, coupled with significant boosts in both stability and cost-efficiency.

Hy3 leverages a Mixture of Experts (MoE) framework, featuring a total of 295 billion parameters, with 21 billion actively engaged parameters. It supports a context length of up to 256K, marking notable advancements in productivity-oriented tasks, such as software development and office automation. Through rigorous expert blind testing, Hy3 has proven to outperform even GLM5.1 in certain aspects.

Presently, Hy3 has been seamlessly integrated into a suite of services, including WorkBuddy/CodeBuddy, Yuanbao, Marvis, and ima, thereby substantially elevating the user experience in practical applications. For instance, the success rate of tasks within WorkBuddy has soared to 90%, while the hallucination rate in Yuanbao has plummeted by over 50%.

The API for Hy3 is now accessible on Tencent Cloud TokenHub, with international API platforms slated for integration in the near future. Its pricing scheme is as follows: 1 yuan per million tokens for input, 4 yuan per million tokens for output, and a mere 0.25 yuan per million tokens for input when leveraging the cache. Furthermore, Hy3 is released under the Apache 2.0 open-source license, empowering developers globally to download and utilize it for commercial applications without incurring any costs. It has also been incorporated into platforms like Huggingface and Modelscope, facilitating broader accessibility and adoption.