Skip to content
by InfinyAI ES

IA en un minutoNewsModels

Models

Tencent launches Hy4 Flash, a cheaper version of its Hy4 model for office work

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

Tencent is rolling out Hy4 Flash, a trimmed-down edition of its Hy4 model meant for everyday office chores such as documents, slides or operating the computer. Hy4 Flash switches on 12 billion of its 224 billion parameters for each token and accepts contexts of one million tokens. WorkIndex, the company's own test, puts its cost per task at one twentieth of Hy4's.

Tencent on Monday, November 16, presented Hy4 Flash, the budget version of Hy4, the flagship model it launched on November 4, and describes it as its most cost-effective Hy model so far. It is the third release in the family, after Hy4 Preview on August 28, according to Dataconomy. The company recommends it for most office tasks, such as in-depth research, handling documents, building slide decks, putting together simple web demos or operating the computer, and keeps the full Hy4 for complex jobs such as long programming projects.

It has 224 billion parameters in total (64 billion of them of a kind Tencent calls Engram), but activates only 12 billion for each token. As with Hy4, it handles images natively and supports a context of up to one million tokens, meaning a great deal of text at once. Tencent says it generates more than 150 tokens per second.

In dollars, its API costs $0.07 per million input tokens and $0.281 per million output tokens, against $0.834 and $2.501 for Hy4. In mainland China, according to Dataconomy, the rate is 0.5 yuan per million input tokens and 2 yuan per million output tokens.

The performance figures come from Tencent, which cites no independent evaluations. On WorkIndex, an internal test built on typical WorkBuddy tasks, the average cost per task is one twentieth of Hy4's and one thirteenth of Hy4 Preview's. A blind evaluation by 85 WorkBuddy users inside the company, covering 259 tasks in 27 categories, put user satisfaction close to that of Claude Haiku 5.5 (at its high setting, according to Dataconomy), with the shortest turnaround time and a cost per task markedly lower than that of any other model tested.

It is already available on Tencent Cloud TokenHub, its platform for API access, and in several of its apps, such as QQ, WeCom, Yuanbao and WorkBuddy, and it will reach OpenRouter soon. Until November 30 it is free in WorkBuddy, and the API carries a 50% discount on TokenHub. Tencent plans to release the model with open weights, without giving a date.

Why it matters · analysis and opinion

Hy4 Flash targets the least glamorous and most used part of AI: everyday work such as summarizing a report, redoing a presentation or fixing a spreadsheet. Tencent leaves long programming to its large model and competes here on price. When you pay by usage, that difference may weigh more than a few points of quality, but every comparison is signed by the company itself and was run with its own internal users. The sensible course is to try it on real tasks while the free promotion lasts and measure how much has to be corrected by hand before moving it into production. The release of the weights, still without a date, is what would allow it to run on in-house servers.

Official source: Tencent · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.