Tencent Hunyuan launches Hy4 flash, with per-task cost just 1/20 of Hy4

TencentHy's team has released Hy4 flash, calling it the most cost-effective Hunyuan model so far. It has 224B total parameters, including 64B Engram parameters, and activates only 12B per token. It takes multimodal input and handles a context of 1 million tokens, with an API price far below Hy4.

The lower cost comes from tuning pre-training, post-training and inference together. Pre-training adds hybrid attention, LatentMoE and Engram, which cuts the KV cache to 1/24 of Hy3 and 1/6 of Hy4. Post-training uses self-distillation and reinforcement learning to shorten output length by 50%. At inference, megakernel and 4bit quantization push decode speed past 150 tokens per second.

On the official WorkIndex benchmark, one task costs 1/13 of Hy4 preview and 1/20 of Hy4. In a blind test with 85 internal users across 259 daily tasks, satisfaction came close to Claude Haiku 5.5, with the lowest cost and the shortest task time among all models compared.

The model is now live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ and ima. Tencent also plans to open source it and keep iterating fast.

SponsoredDomoAI
AI Tools Mentioned

Comments

Loading comments…