Tencent Hunyuan releases the Hy4 flash model, focused on cost-effectiveness for everyday work
Tencent Hunyuan has launched Hy4 flash, with 224B total parameters (including 64B Engram parameters) and 12B activated, supporting multimodality and a 1M context; API pricing is 0.5 yuan for input and 2 yuan for output per million tokens, and 0.02 yuan for cached input hits. The company says it is slightly better than Hy4 preview in overall performance, and on the internal evaluation set WorkIndex its per-task cost is 1/13 that of Hy4 preview and 1/20 that of Hy4. The model is already live in products including Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, will soon be integrated into OpenRouter, and is planned to be open-sourced.
Why read this
AI score 74
The official announcement provides Hy4 flash's pricing, architectural trade-offs, and internal evaluation criteria, which can be used to decide which tier of model to choose for everyday tasks.
Topics
Body · AI translation
@@@space height = 16 @@@
@@@buttonGroupCard
[[buttons]] text = "Tencent Cloud" icon = "tencentcloud" url = "https://cloud.tencent.com/tokenhub/hy4-flash"
@@@
@@@space height =32 @@@ We are launching Hy4 flash, the most cost-effective Hy model to date. It has 224B total parameters (including 64B Engram parameters), 12B activated parameters, supports multimodality and a 1M context length, and its API price is significantly lower than that of Hy4.
@@@space height = 32 @@@
::: center
|API price (per million Tokens)|Input (cache hit) | Input | Output | | ------ | ------ | ------ | ------ | | **Hy4 flash** | **0.02 yuan** | **0.5 yuan**| **2 yuan** | | Hy4 | 0.3 yuan | 6 yuan| 18 yuan |
:::
@@@space height = 32 @@@
Through a series of efforts, we have maximized cost optimization while balancing intelligence and practicality:
- In the pre-training stage, Hy4 flash incorporates architectural features such as hybrid attention, LatentMoE, and Engram. Its KV cache size is only 1/24 that of Hy3 and 1/6 that of Hy4, and extensive scaling ladder experiments ensure that performance meets expectations;
- In the post-training stage, we obtained strong overall intelligence through self-distillation from Hy4, and then used reinforcement learning to reduce the end-to-end task output length by 50%;
- In the evaluation stage, we built WorkIndex, an internal evaluation set based on typical WorkBuddy tasks, and combined it with human evaluation, focusing on the quality of deliverables in daily work, single-turn output length, and end-to-end time, rather than public benchmark scores that are far removed from daily work;
- In the inference stage, we combined megakernel, 4-bit quantization, DFly, and other methods for extreme optimization, enabling Hy4 flash to achieve a decoding speed of over 150 tps and to be deployed and operated at large scale and low cost.
**Ultimately, Hy4 flash is slightly better than Hy4 preview in overall performance; on WorkIndex, its per-task cost is only 1/13 that of Hy4 preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1017 (max tier, busy-hour price).**
@@@space height = 32 @@@
@@@tabContentCard [[tabs]] label = "WorkIndex (internal)" name = "TranslationEvaluation"
[[tabs]] label = "Open Benchmarks" name = "TranslationEvaluation"
To measure the real-world experience, we invited 85 internal WorkBuddy users to conduct blind tests on 259 daily work tasks across 27 categories. The results show that Hy4 flash's user satisfaction is close to that of Claude Haiku 5.5, and its per-task cost is significantly lower than that of all other models, with the shortest task duration as well.
Hy4 flash has been launched on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis, and other products, and will soon be integrated into OpenRouter. Before November 30, users can use it for free for a limited time on WorkBuddy, and the Tencent Cloud TokenHub API is available at a limited-time 50% discount. We recommend using Hy4 flash for most daily tasks such as deep search, document processing, PPT generation, simple front-end display, and computer use, and using Hy4, which has a higher intelligence ceiling, for complex tasks such as long-horizon software engineering. We plan to open-source Hy4 flash and iterate frequently to continuously improve the practicality and efficiency of real work.
@@@testimonialSlider ariaLabel = "Product feedback" itemsPerPage = 2 mobileItemsPerPage = 1 loop = false defaultActivePage = 0 showCount = true countSeparator = " / " countMinDigits = 02 countLabelTemplate = "Page {current} of {total}" emptyText = "No testimonials available" contentEmptyText = "No testimonials available" previousLabel = "Previous testimonial" nextLabel = "Next testimonial"
[[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026111601215756_2ca7df5c2fe9a52ad7abb7456d9f73db.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1794763319;2110123319&q-key-time=1794763319;2110123319&q-header-list=&q-url-param-list=&q-signature=62cad7ed7ed70f0675b80fa8e23a71550f9c6889" brandName = "WeChat Pay" content = "In the evaluation of AI operating a real-device payment client scenario, we evaluated four Agent processes: execution path, knowledge utilization, acceptance checks, and account creation accuracy. Compared with DeepSeek-V4.1-Flash-1017, Hy4 flash was equal or higher in average score across all of them, with about 6% fewer execution steps and about 29% shorter single output length. The execution path is cleaner, the understanding of the injected knowledge base is clearer, and usage is more accurate. In the evaluation of the WeChat Pay code static scanning scenario, compared with DeepSeek-V4.1-Flash-1017, Hy4 flash reduced the time required to complete the scan by 37% and Token consumption by 46%, with comparable reasoning performance; the model's replies are more concise and better able to find the basis for decisions from the context." authorName = "WeChat Pay" authorTitle = "Information Intelligence Lab yilin Security Engineer Simba"
[[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403023983_8ca88aeaa8e8a39d2774d600860162f9.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732561;2109092561&q-key-time=1793732561;2109092561&q-header-list=&q-url-param-list=&q-signature=29ee4e6413e3333f9e0354abb9a212bc37d68856" brandName = "QQ" content = "After integrating Hy4 flash, Xiao Q's multiple Agent capabilities have been further improved: it both answers well and responds quickly. Evaluations show that, compared with DeepSeek-V4.1-Flash, Hy4 flash performs better in music recommendation, channel interaction, Qzone, and long-form horoscope analysis, with the proportion of results meeting requirements higher by 6.16, 2.84, 0.92, and 2.22 percentage points, respectively; in chat companionship, music recommendation, and long-form horoscope analysis scenarios, completion time is shortened by about 11%–27%, enabling users to obtain companionship and help that fit their needs more quickly. In the UGC interaction scenario of the channel assistant, Hy4 flash can stably complete intent recognition and tool invocation under near-real-time constraints. Compared with DeepSeek-V4.1-Flash, the accuracy of judging user intent is on par, the P90 latency of comments is shortened by 39%, peak latency is shortened by 47%, output tokens are reduced by 30%, and the cache hit rate is greatly improved, fully meeting business requirements." authorName = "QQ" authorTitle = "Xiao Q Technical Lead Red Channel R&D Look"