Hy4 preview's parameter count jumped from 295B to 770B—Tencent is officially going all-in on flagship models. It went live for just 3 days and already got overwhelmed with traffic, causing queues to use it. How does Hy4 perform when paired with workbuddy? For the Hy3 generation, Tencent took the small-scale, high-efficiency route. It had only 295B total parameters and 21B active parameters, often compared to MiniMax M3 in external benchmarks. Back then, Tencent proved that even with around 300 billion parameters, a model could still rank in the top tier of domestic AI. Hy4 preview takes a different approach. Total parameters were bumped up to 770B, with 49B active per token, and context length increased from 256K to 1M. For comparison, Zhipu's GLM-5.3 uses the GLM-5 series base with 744B total parameters and 40B active parameters. The two are now in a very similar parameter range. This time, Tencent has released a model truly designed to flagship specs, with performance better than GLM 5.2 but slightly behind GLM 5.3. Tencent conducted blind tests internally. Hy4 preview scored an average of 2.99, while GLM-5.3 scored 2.92. Hy4's win/draw/loss rates against GLM-5.3 were 46.8%, 12.8%, and 40.4%, respectively. Based on real-world experience, I’d interpret this as Hy4 having entered the flagship competition zone of GLM-5.3, but it needs to be paired with workbuddy to unleash its full potential.