Tencent Open-Sources 770B-Parameter Hy4 Preview With 1M-Token Context
Updated
Updated · tencent.com · Aug 28
Tencent Open-Sources 770B-Parameter Hy4 Preview With 1M-Token Context
3 articles · Updated · tencent.com · Aug 28
Summary
Hy4 preview is now live through Tencent’s WorkBuddy, CodeBuddy, Yuanbao and ima, with free access on WorkBuddy and CodeBuddy for two weeks.
770 billion total parameters, 49 billion active parameters and a context window above 1 million tokens position the model for coding, office work and scientific research workloads.
163 experts testing 203 engineering tasks gave Hy4 preview 2.99 out of 4, narrowly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94 in Tencent’s internal blind evaluation.
31.8% higher end-to-end throughput came after the model helped optimize its own training methods and inference system, an early recursive self-improvement loop Tencent highlighted.
API pricing starts at $0.834 per million input tokens, with Tencent saying more Hy4-series models will follow under its preview-first rollout strategy.
Tencent's new 770B AI optimized its own code to run 31% faster; are we witnessing the dawn of truly self-improving algorithms?
Why would a tech giant give away a massive 770-billion parameter AI model for free under an open-source Apache 2.0 license?
How will independent developers manage the massive hardware demands of deploying this 770-billion parameter AI, even with its efficient MoE architecture?