Tencent Hunyuan Cuts Hy4 flash Input Price to 0.5 Yuan

Tencent Hunyuan Cuts Hy4 flash Input Price to 0.5 Yuan

Tencent Hunyuan released Hy4 flash on November 16, positioning it as the best value model in the Hy series. It is already live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom and QQ. Tencent says it will soon be available on OpenRouter and will be open-sourced later.

The launch comes only about ten days after the official Hy4 release. Hy4 took the capability route; Hy4 flash shifts the goal to everyday office work: deep search, document handling, building slide decks and simple front-end pages. Tencent's own advice is blunt: for complex tasks such as long-horizon software engineering, Hy4 is still the recommended model.

How the model was slimmed down

Hy4 flash has 224B total parameters, of which 64B are Engram parameters, a scalable lookup-style module. Only 12B are activated per token. The model accepts multimodal input and supports a 1M-token context window.

Tencent breaks the cost cutting into three stages. In pre-training it used hybrid attention, LatentMoE and Engram, so the KV cache is just 1/24 of Hy3's and 1/6 of Hy4's, which eases memory pressure considerably at long context lengths. In post-training it first self-distilled from Hy4 to preserve overall capability, then used reinforcement learning to cut the output length of end-to-end tasks in half. At inference time it stacks a megakernel, 4-bit quantization and DFly, pushing decoding speed above 150 tokens per second.

The halved output length matters most to users billed per token. If a model writes half as much for the same task, the bill shrinks by half, and that saving is separate from whatever the unit price drops by.

Price comparison

Hy4 flash costs 0.5 yuan per million input tokens, 2 yuan per million output tokens and 0.02 yuan for cache hits. According to The Paper, the official Hy4 released on November 4 is priced on the same basis at 6 yuan for input, 18 yuan for output and 0.3 yuan for cache hits. By rough calculation, input is 12 times cheaper, output 9 times cheaper and cache hits 15 times cheaper.

Beyond unit prices, Tencent used its in-house WorkIndex benchmark to calculate per-task cost: Hy4 flash costs 1/13 as much as Hy4 preview and 1/20 as much as Hy4, and only 1/8 of DeepSeek-V4.1-Flash (max setting, peak-hour pricing). WorkIndex is made up of typical office tasks from WorkBuddy, including building PPT decks, editing Word files and fixing Excel sheets. Tencent says Hy4 flash scores slightly higher overall than Hy4 preview on this set.

Compare that with Qwen4-Max, which launched today: it charges 1.50 US dollars per million input tokens, or roughly 10.8 yuan at an assumed rate of 7.2 yuan to the dollar, more than twenty times Hy4 flash. The two target different jobs, one chasing leaderboards and the other handling everyday chores, so the multiple mainly shows how wide the price range of Chinese models has become.

ModelInput (yuan / million tokens)Output (yuan / million tokens)Cache hit
Hy4 flash0.520.02
Hy4 (official release)6180.3

Internal blind test and what remains undisclosed

Tencent recruited 85 internal WorkBuddy users for a blind test covering 27 categories and 259 everyday work tasks. By Tencent's account, Hy4 flash's satisfaction score came close to Claude Haiku 5.5, with shorter task times than the other models tested and the lowest per-task cost.

All of these conclusions come from Tencent's internal evaluation. The WorkIndex questions are not public, and the testers were Tencent employees. No third-party organization has yet published independent results for Hy4 flash, and the open-source release date and license have not been announced. Until then, the easiest way for developers in China to check the claims is to run their own real tasks through the model before the half-price offer ends at the end of the month.

Sources: Tencent Hunyuan official announcement, CNMO, CocoLoop, The Paper, DataLearner; input, output and cache-hit prices were checked against Tencent Cloud TokenHub's price list, and Hy4 pricing follows The Paper's figures.