Tencent Launches Hy4 Flash to Deliver Cost-Efficient AI for Everyday Workflows
· GlobeWire Daily Editorial Desk · Source: www.prnewswire.com

Image: Tencent · Photo source · Resized and letterboxed to 16:9; no AI generation
Tencent announced the release of Hy4 flash, a new addition to its Hy model family designed to support high-speed, cost-efficient artificial intelligence deployment for everyday workflows. According to the company announcement, the model is now available across several platforms, with promotional pricing and free access options offered through November 30.
Model Architecture and Performance Specs
Hy4 flash features 224 billion total parameters, which includes 64 billion Engram parameters, alongside 12 billion active parameters. It supports multimodal inputs and a context length of up to one million tokens. Tencent states that the model is priced at approximately US$0.070 per million input tokens, US$0.281 per million output tokens, and US$0.003 per million cached input tokens.
Key architectural and engineering highlights reported by Tencent include:
- Pre-training: Combines hybrid attention, LatentMoE, and Engram features, resulting in a KV cache size reduced to 1/24 of Hy3 and 1/6 of Hy4.
- Post-training: Incorporates reinforcement learning to reduce average end-to-end task output tokens by 50 percent.
- Inference Optimization: Uses megakernel, 4-bit quantization, and DFly techniques to achieve decoding speeds exceeding 150 tokens per second.
Evaluation and Real-World Utility
Instead of relying solely on public benchmarks, Tencent reported using WorkIndex—an internal benchmark based on WorkBuddy tasks—alongside human evaluations. According to internal data cited in the release:
- Hy4 flash slightly outperforms the larger Hy4 Preview in overall effectiveness while operating at a lower cost.
- In a blind evaluation involving 85 internal users across 259 tasks, the model achieved user satisfaction comparable to Claude Haiku 5.5, with low per-task costs and fast completion times.
- The model is recommended for everyday tasks such as deep research, document processing, PowerPoint generation, and simple front-end displays, whereas the larger Hy4 model is suggested for complex tasks like long-horizon software engineering.
Availability and Practical Implications
Hy4 flash is currently available across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with an OpenRouter integration planned. Users can access the model for free on WorkBuddy and receive a 50 percent discount on standard pricing via Tencent Cloud TokenHub through November 30. Tencent also stated plans to open-source Hy4 flash in future iterations.
Readers can verify these announcements by checking official Tencent press releases, visiting the Tencent Cloud TokenHub portal, or reviewing platform update logs for WorkBuddy and Yuanbao.
AI-assisted reporting; source claims are attributed and have not been independently verified.



