Tencent Hunyuan Releases Hy4 Flash at One Twentieth the Cost of Hy4

By 褚杏娟Original title: 腾讯混元发布Hy4 flash,单任务成本仅为 Hy4 的 5%
78AI Score

Tencent Hunyuan released Hy4 Flash, a 224B-parameter (12B activated) multimodal model with 1M context, priced at RMB 0.5 per million input tokens. Through LatentMoE, Engram, 4-bit quantization and megakernel optimizations, its KV cache is one sixth of Hy4's and per-task cost drops to one twentieth, while slightly outperforming the 770B Hy4 preview. Hunyuan plans to open-source it with frequent iterations.

Full Report

On November 16, Tencent Hunyuan released its latest language model, Hy4 flash. The model has a total parameter count of 224B (including 64B Engram parameters), with only 12B activated parameters, and supports multimodality and a 1M context length. Hy4 flash is the most cost-effective Hy model to date: RMB 0.5 per million input tokens and RMB 2 per million output tokens, and only RMB 0.02 when cache hits occur, a substantial price reduction compared with Hy4.

To achieve extreme cost-effectiveness while balancing intelligence level and practicality, the Hunyuan team carried out a series of efforts:

  • In the pretraining stage, Hy4 flash incorporated architectural features such as hybrid attention, LatentMoE, and Engram. The KV cache size is only 1/24 that of Hy3 and 1/6 that of Hy4, and extensive scaling ladder experiments were conducted to ensure the results met expectations.
  • In the post-training stage, strong overall intelligence was obtained through self-distillation from Hy4, and reinforcement learning then reduced the end-to-end task output length by 50%.
  • In the evaluation stage, an internal evaluation set called WorkIndex was built based on typical WorkBuddy tasks, and combined with human evaluation, it focuses on the quality of deliverables in daily work, single-turn output length, and end-to-end time, rather than public leaderboard scores that are far removed from daily work.
  • In the inference stage, extreme optimization was carried out by combining megakernel, 4-bit quantization, DFly, and other methods, 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 the larger Hy4 preview (770B total parameters, 49B activated parameters) in overall performance, but at significantly lower cost. Based on comprehensive measurements using the internal evaluation set WorkIndex (covering typical WorkBuddy tasks such as PPT creation and modification, document editing, spreadsheet repair, audio and video processing, and visual understanding), the per-task cost of Hy4 flash is only 1/13 that of Hy4 preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1027 (Max tier, peak-hour pricing).

Blind tests with 85 internal WorkBuddy users at Tencent showed that across 259 tasks in 27 categories, user satisfaction with Hy4 flash was close to that of Claude Haiku 5.5, while its per-task cost was significantly lower than that of all other models, and its task duration was also the shortest.

Currently, 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, users can use it free for a limited time on WorkBuddy, and the Tencent Cloud TokenHub API is available at a limited-time 50% discount.

According to the official statement, Tencent Hunyuan always takes "model practicality" as its core principle, pursuing extreme cost-effectiveness while continuously pushing the upper limit of intelligence.

Hy4, released on November 4, entered the ranks of the strongest open-source models for the first time, ranking second among Chinese models on the intelligence index of the internationally authoritative model testing leaderboard Artificial Analysis, and first among open-source models on the Agent Arena leaderboard. The release of Hy4 flash this time achieves a better balance between intelligence and efficiency in real productivity tasks.

The Hunyuan team stated that users are recommended to use Hy4 flash for most daily tasks such as deep search, document processing, PPT generation, simple front-end display, and Computer Use, and to use Hy4, which has a higher intelligence ceiling, for complex tasks such as long-horizon software engineering. At the same time, Hunyuan plans to open-source Hy4 flash and iterate frequently to continuously improve the practicality and efficiency of real work.