Lightweight version 12 days after the flagship's official release

China's Tencent on Nov. 16 (the afternoon of the same day Japan time) released "Hy4 flash," a lightweight version of its "Hunyuan" AI models, on an official announcement page on Hunyuan's research site. The company described it as the most cost-effective Hy model to date. It has 224 billion total parameters (64 billion of them in a part it calls "Engram"), and only 12 billion are active in each pass. It is multimodal, handling images and other inputs in addition to text, and can read 1 million tokens at once.

On Oct. 29, Tencent said it would soon release the official version of its flagship "Hy4" along with Hy4 flash. The official version of Hy4 came out on Nov. 4, but Hy4 flash was not released at that time (ORIENS's earlier report). The announcement page recommends using Hy4 flash for daily work and the higher-ceiling Hy4 for complex tasks such as long software development.

API priced at 0.5 yuan input, 2 yuan output

According to the announcement page, API pricing is 0.5 yuan per million input tokens and 2 yuan per million output tokens. Input that hits the cache, when the opening part of the same input is reused, costs 0.02 yuan. The flagship Hy4 costs 6 yuan for input, 18 yuan for output and 0.3 yuan for cached input, and Tencent said the price was sharply lower than Hy4's. Calculated from the figures in the official table, input costs 1/12 and output 1/9 as much.

API prices of Hy4 flash and Hy4 (per million tokens, yuan). From Tencent's official announcement page
  • Hy4 flash
  • Hy4
0.020.3
Input (cache hit)
0.56
Input
218
Output

Cost per task 1/20 of Hy4 in internal evaluation, slight decline in long coding tasks

Tencent said it evaluated the model with "WorkIndex," an internal evaluation set built from typical tasks in its business AI agent "WorkBuddy," along with human evaluation. WorkIndex consists of eight internal evaluation sets, including creating and editing slides, editing documents, fixing spreadsheets, processing audio and video, and understanding images. The company said it emphasized the quality of deliverables in daily work, the length of each output and the time to finish a task over scores on public benchmarks.

According to the company, Hy4 flash slightly outperformed August's "Hy4 preview" in overall performance. Its cost per task on WorkIndex was 1/13 that of Hy4 preview and 1/20 that of Hy4. It said the cost was 1/8 that of Chinese company DeepSeek's "DeepSeek-V4.1-Flash-1017" when that model was run at its maximum reasoning depth and at peak-hour prices.

Tencent also ran a blind test in which 85 internal WorkBuddy users compared models on 259 daily tasks across 27 categories without knowing which model was which. According to Tencent, Hy4 flash's satisfaction rating was close to that of U.S. company Anthropic's "Claude Haiku 5.5," its cost per task was far lower than all other models, and it took the least time. All of these are Tencent's own evaluations, and public benchmark results are shown only in a chart in the appendix.

In an evaluation by a WorkBuddy product manager posted on the announcement page, average task time was 70% shorter than with Hy4 and the average cost was about 1/15. Performance was about the same on daily tasks, while it declined slightly on long-running coding tasks.

Memory footprint cut to 1/6 of Hy4, output length halved

Tencent said that in pre-training, the first stage of training, it adopted architectures such as "hybrid attention," which combines several attention mechanisms, "LatentMoE" and "Engram." The size of the intermediate data the model keeps while reading through text (the KV cache) is 1/24 that of the previous-generation Hy3 and 1/6 that of Hy4, it said.

In post-training, it raised overall capability through "self-distillation," using Hy4's outputs as a model, and used reinforcement learning to cut the length of output needed to finish a task end to end by 50%. At the serving stage, it said methods such as a "megakernel" that speeds up processing by combining it, 4-bit quantization and "DFly" let the model generate text at more than 150 tokens per second.

Available in Tencent services, free on WorkBuddy until Nov. 30

Hy4 flash is available through the API of "TokenHub," the model platform of Tencent's cloud, as well as in WorkBuddy, the AI assistant "Yuanbao," the enterprise chat app "WeCom," "QQ," the knowledge management tool "ima" and "Marvis." Tencent said it will soon be available on U.S.-based OpenRouter, which offers models from multiple companies through a single API.

Until Nov. 30, it can be used for free on WorkBuddy, and the TokenHub API is offered at half price. Tencent said it plans to release the weights of Hy4 flash and update it frequently, but did not give a timeline for the release.