ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Hy4 Flash Launches: Tencent Hunyuan's 224B Lightweight Model Cuts Per-Task Cost to One-Twentieth of Hy4

Hy4 Flash Launches: Tencent Hunyuan's 224B Lightweight Model Cuts Per-Task Cost to One-Twentieth of Hy4

AI information • Admin • • 26 views

On November 16, 2026, Tencent Hunyuan officially launched Hy4 flash, the first model in the Hy series sold on being lightweight and cheap: 224B total parameters with only 12B activated per inference, yet it keeps the flagship Hy4's multimodal input and 1-million-token context window, with list API prices cut to RMB 0.5 per million input tokens and RMB 2 per million output tokens.

Two-thirds of the parameters are gone, and the savings come from more than one place

Hy4 flash has roughly a third of the flagship Hy4's 770B total parameters, and activated parameters drop from 49B to 12B. Of the total, 64B are Engram parameters — a conditional memory technique that stores common text patterns in a fast lookup table, so the model retrieves results instead of recomputing them every time. Pre-training combines hybrid attention, LatentMoE and Engram, while inference uses megakernel, 4-bit quantization and other techniques to push decoding past 150 tokens per second. The KV cache is just 1/24 of Hy3's and 1/6 of Hy4's, which brings memory usage down with it — that is what makes the RMB 0.5 and RMB 2 pricing possible. Against Hy4's list prices, input, output and cache hits are about 12x, 9x and 15x cheaper, although Hy4 currently has a separate 50% discount, so the real gap is smaller than the list-price comparison suggests.

How much cheaper, and how much performance is lost

Tencent published a comparison on WorkIndex, its own benchmark built from typical WorkBuddy tasks: collaborative PPT editing fell from 78.9 with Hy4 to 74.0, while the cost per task dropped from RMB 1.714 to RMB 0.105; automated tasks fell from 56.2 to 52.2, with the cost per task down from RMB 0.908 to RMB 0.038. Overall, the cost per task is roughly one-twentieth of Hy4's and one-eighth of DeepSeek V4.1-Flash's. In a blind test, 85 WorkBuddy users rated 259 tasks, and satisfaction came close to Claude Haiku 5.5. One caveat: all of these numbers come from Tencent's own internal evaluation, WorkIndex is not public, and no independent verification has appeared yet.

Everyday office work goes to flash, complex development still needs the flagship

Tencent's division of labor is straightforward: deep research, document processing, slide creation, spreadsheet editing, simple front-end web development and computer use — everyday tasks like these go to Hy4 flash, while long software development workflows and other complex work stay with Hy4. The model is already live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter to follow. Until November 30, it is free to use in WorkBuddy and half price on TokenHub's API. Anyone curious can run a few real tasks in WorkBuddy first and then decide whether to build it into their workflow — a more reliable test than the price list alone. For the flagship's capabilities, see our earlier coverage: Tencent Hunyuan Hy4 Officially Launches: First With Vision and Computer Use, Weights to Be Open-Sourced.

Recommended Tools

More