Tencent Hunyuan released Hy4flash today. The official description for this model is quite straightforward: it is the most cost-effective Hy model to date, specifically designed for everyday tasks.
The specifications indicate that the total number of parameters is 224B (of which 64B are dedicated to Engram-related parameters), while the actual number of active parameters amounts to only 12B. This model supports multi-modal processing and can handle up to 1 million context elements, enabling a decoding speed of over 150 tokens per second.
Nearly every aspect of the process has been optimized to save costs. For pre-training, hybrid attention mechanisms, LatentMoE, and Engram architectures were employed; the size of the KV cache was reduced to only 1/24 of that used in Hy3. During post-training, data was distilled using Hy4 techniques, and reinforcement learning was applied to halve the length of the output generated by the model. On the inference side, 4-bit quantization and megakernel technologies were implemented.
The cost figures were calculated using Tencent’s self-developed WorkIndex evaluation dataset: the cost per task is 1/13 of that of Hy4preview and 1/20 of that of Hy4; it is also 1/8 of the cost associated with DeepSeek-V4.1-Flash-1017 (at its maximum configuration and during peak usage times). 85 internal users of WorkBuddy participated in a blind test involving 259 routine tasks, and they reported satisfaction levels comparable to those achieved by Claude Haiku 5.5 – moreover, WorkBuddy took the shortest amount of time to complete these tasks.
These are all evaluation datasets and user groups owned by Tencent itself, so it’s necessary to apply a certain degree of skepticism when referring to them. However, the approach is clear: instead of making the scores on these lists public, the focus should be solely on “how much money it costs and how long it takes to complete the task”.
Tencent has not claimed that it is an all-round solution either: Its performance in long-distance encoding tasks has decreased slightly compared to Hy4. The company recommends using Hy4flash for tasks such as advanced searches, document processing, creating PPTs, and basic computer operations; however, more complex software development projects should still be handled by Hy4.
It has already been integrated with services such as Tencent Cloud TokenHub, WorkBuddy, Tencent Yuanbao, WeCom, QQ, and iMa, and will soon be connected to OpenRouter as well. Until November 30th, WorkBuddy offers a limited-time free trial, while the TokenHub API is available at a 50% discount. The development team has stated that they plan to open-source these tools in the future.