Tencent launches Hunyuan Turbo on Tencent Cloud

Background

Hunyuan Turbo was a step in the Hunyuan line, and Tencent compared its efficiency claims against earlier Hunyuan models ^[1]. The launch was aimed at enterprises and developers who wanted to use a large language model through a cloud service.

Cloud delivery means the model runs on the provider's servers, and customers pay for the requests they send. Efficiency claims for models like this are usually framed as training efficiency, inference efficiency and decoding speed, which are the measures this announcement uses. Decoding speed is how quickly the model writes out its output.

What happened

Tencent introduced its next-generation Hunyuan Turbo model, reported on 5 September 2024, and made it available on Tencent Cloud ^[1]. The launch was announced by Qiu Yuepeng, Tencent vice-president and chief operating officer of the Cloud and Smart Industries Group ^[1]. Enterprises and developers could integrate the model directly through Tencent Cloud ^[1].

Key details

Tencent claimed training efficiency up 108%, inference efficiency up 100% and decoding speed up 20% compared with earlier Hunyuan models ^[1]. Inference cost was cut by 50%, and input and output prices were halved against the predecessor ^[1]. The company said the model matched GPT-4o on various benchmarks ^[1].

These are Tencent's own figures. The source reports them as claims and does not show independent test results.

Impact

For developers, the most practical change described was cost. Halving the price and the inference cost means that an application built on the model would cost less to run at the same volume ^[1]. Access through Tencent Cloud also meant that the model was offered as a cloud service rather than as a download.

A claim of matching a rival model on benchmarks shows the result of those particular tests, so independent comparisons remain necessary for a fuller picture.