AI News

Tencent Hunyuan releases the Hy4 flash model: 224B total parameters, 12B activated, with a per-task cost about 5% of Hy4

IT Home · Artificial IntelligenceUpdated Nov 17 00:53

AI summary

The Tencent Hunyuan team has released Hy4 flash, with 224B total parameters (including 64B Engram parameters), 12B activated per token, and support for multimodal input and a 1 million token context. According to the official statement, on the WorkIndex evaluation set its per-task cost is 1/13 that of Hy4 preview and 1/20 that of Hy4, with decoding speed exceeding 150 tps. The model has already launched on products including Tencent Cloud TokenHub, WorkBuddy, and Yuanbao, and is planned to be open-sourced.

Why read this

SelectedAI score 81

The original text provides the total parameter count, activated parameters, and per-task cost comparison, so readers can use these to judge the trade-offs of this tier of model in cost-sensitive tasks.

Topics

Tags

Body · AI translation

Tencent Hunyuan Hy4 flash Model Released: 224B Total Parameters, 12B Activated, Cost per Task About 5% of Hy4

2026-11-16 13:09 IT Home - Wenzhou

Thanks to IT Home user Xinghezhimeng for the tip submission!

IT Home, November 16 news: The Tencent Hunyuan team today announced the official launch of Hy4 flash, which the company positions as the most cost-effective Hy model to date.

The model has 224B total parameters, including 64B Engram parameters, with 12B parameters activated per token. It supports multimodal input and a 1 million token context length, and its API price is significantly lower than that of Hy4.

Tencent said that the high cost-effectiveness of Hy4 flash comes from coordinated optimization across three stages: pre-training, post-training, and inference. In the pre-training stage, architectures such as hybrid attention, LatentMoE, and Engram (IT Home note: Engram is a scalable lookup-table-style parameter module used to expand model capacity at lower compute cost) were introduced, reducing the KV cache size to 1/24 of Hy3 and 1/6 of Hy4. In the post-training stage, comprehensive intelligence is obtained through self-distillation from Hy4, and reinforcement learning is then used to compress the end-to-end task output length by 50%. In the inference stage, methods such as megakernel, 4-bit quantization, and dflash are combined, achieving a decoding speed of over 150 tps and supporting large-scale, low-cost deployment and operation.

On WorkIndex, the official evaluation set of typical WorkBuddy tasks, the cost per task of Hy4 flash is only 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (max tier, peak-hour pricing).

In terms of overall performance, Hy4 flash is slightly better than Hy4 preview. This comparison framework focuses on the output quality of daily work, single-round output length, and end-to-end time, rather than public leaderboard scores.

To measure the real user experience, Tencent invited 85 internal WorkBuddy users to conduct blind tests on 259 daily work tasks across 27 categories.

The results show that user satisfaction with Hy4 flash is close to that of Claude Haiku 5.5, its cost per task is significantly lower than all models in the comparison, and its task duration is also the shortest.

Hy4 flash has been launched on products including Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, and will soon be integrated into OpenRouter.

Before November 30, 2026, 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.

Tencent recommends using Hy4 flash for daily tasks such as deep search, document processing, PPT generation, simple front-end display, and Computer Use, while complex tasks such as long-horizon software engineering should be left to Hy4, which has a higher intelligence ceiling. The company also said it plans to open-source Hy4 flash and maintain high-frequency iteration.

Source

IT Home · Artificial Intelligence