Tencent Hunyuan officially launched Hy4flash on November 16. Positioned for everyday work, it is the most cost-effective Hy model to date. The model has 224B total parameters (including 64B Engram parameters) and 12B activated parameters. It supports multimodality and a 1M context. Its API price is significantly lower than that of Hy4. It is already available on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, and will soon be integrated into OpenRouter.

Cost optimization runs through the entire pipeline: pretraining introduces hybrid attention, LatentMoE, and Engram architectures, while KV cache is only 1/24 that of Hy3 and 1/6 that of Hy4; post-training uses self-distillation from Hy4 to achieve a comprehensive intelligence level, and reinforcement learning reduces end-to-end task output length by 50%; inference combines megakernel, 4-bit quantization, and DFly, among other methods, achieving decoding speeds above 150 tps and enabling large-scale, low-cost deployment.
For evaluation, an internal benchmark set, WorkIndex, was built based on typical WorkBuddy tasks and combined with human evaluation. It focuses on output quality, output length, and time taken rather than public leaderboard scores.

It is reported that the model's overall performance is slightly better than Hy4preview. Its per-task cost on WorkIndex is only 1/13 of the former, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (max tier, peak-hour pricing). WorkIndex includes 8 internal task sets and benchmarks against 7 public Agent benchmarks. A blind test by 85 internal WorkBuddy users on 259 everyday tasks across 27 categories showed satisfaction close to Claude Haiku5.5, with per-task cost significantly lower than all other models and the shortest task duration.
It can be used for free for a limited time on WorkBuddy until November 30, and the Tencent Cloud TokenHub API is available at a limited-time 50% discount. The official recommendation is to switch everyday tasks such as deep search, document processing, PPT generation, and Computer Use to this model, while complex tasks such as long-horizon software engineering should still use Hy4, which has a higher intelligence ceiling; internal evaluations also indicate a slight decline in its long-horizon Coding tasks. The team plans to open-source it and iterate frequently. Entering everyday office work with extreme cost-effectiveness continues the trend of large-model competition shifting from capability ceilings to unit cost and delivery efficiency.




