Tencent Hunyuan releases Hy4 flash: the cost of a single task is one-twentieth of Hy4

📅 2026-11-16

Abstract:

Today, Tencent Hunyuan’s latest language model Hy4 flash was released. The total parameter amount of the model is 224B (including 64B Engram parameters), the activation parameters are only 12B, and it supports multi-modality and 1M context length. Hy4 flash is the most cost-effective Hy model to date: input is 0.5 yuan per million Tokens, output is 2 yuan, and only 0.02 yuan is needed when hitting the cache. The price is significantly lower than Hy4.


According to reports, in order to achieve the ultimate cost-effectiveness while taking into account the level of intelligence and practicality, the Hunyuan team has done a series of work:

In the pre-training stage, Hy4 flash adds architectural features such as hybrid attention, LatentMoE, and Engram. The KV cache size is only 1/24 of Hy3 and 1/6 of Hy4, and a large number of scaling ladder experiments ensure that the effect meets expectations.

In the post-training stage, a strong comprehensive intelligence level is obtained through self-distillation Hy4, and then the end-to-end task output length is reduced by 50% through reinforcement learning.

In the evaluation phase, an internal evaluation set WorkIndex based on typical WorkBuddy tasks was constructed, combined with human reviews, to focus on the product effects of daily work, single-round output length, and end-to-end time, rather than public list scores that are far away from daily work.

In the inference stage, combined with megakernel, 4bit quantization, dflash and other means for extreme optimization, the Hy4 flash decoding speed exceeds 150 tps, and can be deployed and operated on a large scale at low cost.

In the end, Hy4 flash is slightly better than Hy4 preview with larger parameters in terms of overall effect (total parameters are 770B, activation parameters are 49B), but the cost is significantly lower. According to comprehensive calculations based on the internal evaluation set WorkIndex (covering typical WorkBuddy tasks such as PPT production and modification, document editing, table repair, audio and video processing, and visual understanding), the single task cost of Hy4 flash is only 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1027 (Max level, busy hour price).

It is reported that Hy4 flash has been launched on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, Enterprise WeChat, QQ, ima, Marvis and other products, and will soon be connected to OpenRouter. Before November 30, users can use WorkBuddy for free for a limited time, and use Tencent Cloud TokenHub API at a 50% discount for a limited time.

Related tags

Related articles

Comments

0/500
Captcha (click to refresh)
No comments yet