Tencent has launched its AI model Hy4 flash, aimed at everyday workflow tasks.
Hy4 flash is the latest model in Tencent's Hy family and is positioned as a lower-cost option for document processing, spreadsheet repair, slide creation and visual understanding. It supports multimodal inputs and a context length of up to 1 million tokens.
Tencent priced Hy4 flash at about USD $0.070 per million input tokens, USD $0.281 per million output tokens and USD $0.003 per million cached input tokens, making it the lowest-cost Hy model so far.
The model has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters. Tencent said Hy4 flash uses LatentMoE and Engram architecture, while its KV cache size is one-sixth that of Hy4 and one twenty-fourth that of Hy3.
Cost focus
Rather than emphasising public benchmark rankings, Tencent assessed the model against what it described as common workplace tasks. Its internal WorkIndex benchmark covered PowerPoint creation and editing, document editing, spreadsheet repair, audio and video processing, and visual understanding.
On that measure, Tencent said Hy4 flash slightly outperformed the larger Hy4 Preview model in overall effectiveness while running at far lower cost. Per-task cost was one-thirteenth of Hy4 Preview, one-twentieth of Hy4 and one-eighth of DeepSeek-V4.1-Flash-1017 at peak-hour Max tier pricing, according to the company.
Tencent also said reinforcement learning cut average end-to-end task output tokens by 50% compared with earlier approaches in the Hy line, contributing to lower operating costs in practical deployment.
Speed claims
According to Tencent, Hy4 flash can decode at more than 150 tokens per second. It attributed that to a mix of inference techniques including megakernel optimisation, 4-bit quantisation and DFly.
The model inherits elements of Hy4's architecture, which Tencent said helped preserve general intelligence despite the lower number of active parameters. The approach reflects a broader trend in AI model development as companies try to reduce serving costs without giving up too much performance on common business tasks.
In a blind internal evaluation involving 85 Tencent WorkBuddy users across 259 tasks in 27 categories, Tencent said Hy4 flash delivered user satisfaction close to Claude Haiku 5.5. It also recorded the lowest per-task cost among the models tested and the shortest completion time, according to the company.
Wider rollout
Hy4 flash is available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. OpenRouter integration is due next, Tencent said.
The launch comes shortly after Tencent introduced Hy4, which it described as one of its leading open-source AI models. Tencent said Hy4 ranked second among Chinese models on the Artificial Analysis Intelligence Index and first among open-source models on the Agent Arena leaderboard.
Hy4 flash is intended for most routine work uses, including deep research, document processing, PowerPoint deck generation, simple front-end displays and computer use. For more complex tasks such as long-horizon software engineering, Tencent said users should turn to Hy4 instead.
Tencent also said it intends to open-source Hy4 flash and update it frequently to improve its practical utility and efficiency for real-world work.