Tencent has released Hy4 flash, a new addition to its Hunyuan model family priced aggressively for everyday tasks. The model has 224 billion total parameters, including 64 billion Engram parameters, but activates only 12 billion per inference. It supports multimodal inputs and a one-million-token context, with decoding speeds above 150 tokens per second.
Its pricing is the sharpest differentiator: 0.5 yuan per million input tokens, 2 yuan per million output tokens and 0.02 yuan for cache hits, compared with 6, 18 and 0.3 yuan respectively for the full Hy4 model. The reductions amount to roughly one twelfth, one ninth and one fifteenth of the original prices.
The cost cuts come from a combination of engineering techniques. Hybrid attention, LatentMoE and the Engram architecture reduce the KV cache to one twenty-fourth of Hy3's and one sixth of Hy4's. Self-distillation and reinforcement learning cut end-to-end output length by half, while megakernels, 4-bit quantization and the DFly optimization speed up inference. In blind tests involving 85 users and 259 tasks, satisfaction approached that of Claude Haiku 5.5.
Hy4 flash is now available on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ and other products, with free access until the end of November and open-source plans announced.





