Overview
- Tencent disclosed Friday that it has published a Hy4 preview as an open-source Mixture-of-Experts model and is rolling the preview into products and APIs including Yuanbao, CodeBuddy, WorkBuddy and ima.
- The company said it is offering limited free access on WorkBuddy and CodeBuddy and that developers can reach Hy4 via Tencent Cloud TokenHub and third-party OpenRouter APIs.
- Tencent published internal figures that describe Hy4 as having 770 billion total parameters with about 49 billion active parameters for a request and a context window exceeding 1 million tokens, but the firm has not released full technical specifications and outside groups have not independently verified these numbers.
- In a blind internal evaluation of 163 experts and 203 engineering tasks, Tencent reported Hy4 scored 2.99 out of 4.00, a slight edge over Z.ai’s GLM-5.3 and Moonshot’s Kimi K3, though those results are company-led rather than third-party benchmarks.
- Hy4 builds on Tencent’s Hy3 MoE lineage and introduces self-optimization steps for training and inference that Tencent says improve coding, data analysis and research workflows; the release sharpens competition among Chinese open-model providers and raises the need for independent testing to confirm real-world gains.