Tencent Hunyuan Hy4 Flash Goes Open Source, with Inference Costs Just 1/20 of Hy4
1 day ago / Read about 0 minute
Author:小编   

Tencent has launched Hunyuan Hy4 Flash, its natively multimodal model built for speed and cost efficiency in everyday office work and Agent-based tasks. The model has 224B parameters, including 64B Engram parameters, activates 12B per token, supports a 1-million-token context window, and is priced far below both its sister models and comparable mainstream Flash-tier offerings. It is now open source. In real-world testing, the model handles tasks such as 3D content generation, mini-game development, financial report analysis, and multi-page PPT creation with ease, delivering output at 200 tokens per second—and it shows none of the intelligence degradation that typically plagues Flash models. Thanks to coordinated optimization across four stages—pre-training, post-training, evaluation, and inference—the cost per task has been cut to 1/20 of the original Hy4, while performance comprehensively surpasses Hy4 Preview. Since launch, call volume has climbed rapidly, topping the daily rankings on major LLM aggregation platforms, and the model is well suited to a wide range of high-frequency everyday productivity tasks.