Hunyuan Turbo

What it is

Hunyuan Turbo is a large language model (LLM) in Tencent's Hunyuan family. Tencent described it as a next-generation Hunyuan language model, and it was reported on 5 September 2024 ^[1]. It is not something the public downloads and runs on its own machines; instead Tencent made it available on Tencent Cloud, so that enterprises and developers could integrate it directly into their own products ^[1]. The announcement came from Qiu Yuepeng, a Tencent vice president and the chief operating officer of the company's Cloud and Smart Industries Group ^[1]. Delivering a model through a cloud platform is a common way for model makers to reach business customers: the customer's software sends text to the model over an interface and gets text back, paying for what it uses rather than maintaining the hardware that runs the model.

Efficiency first

Tencent presented the model mainly as an efficiency upgrade over its predecessors. The company claimed a 108% gain in training efficiency, a 100% gain in inference efficiency and a 20% gain in decoding speed ^[1]. These three measures describe different stages of a model's life. Training efficiency is about how much computing power is needed to reach a given level of quality while the model is being built. Inference efficiency is about how much it costs to answer a request once the model is finished and serving users. Decoding speed is about how fast the model produces its answer, token by token, where a token is a small chunk of text such as a word or part of a word. Improvements on these axes change who can afford to build a model and who can afford to put it into a product.

Price and benchmark claims

Tencent said inference costs were cut by 50%, with input and output prices halved compared with the previous model ^[1]. It also said Hunyuan Turbo matched GPT-4o on various benchmarks ^[1]. Benchmarks are standard test sets used to compare models against one another; a claim of parity with a widely used reference model is a statement about broad capability, not a guarantee for any particular task or language. Pricing matters here because language models are usually billed per token, so halving the per-token price directly changes what an application costs to run at scale.

The successor

In February 2025 Hunyuan Turbo was succeeded by Hunyuan Turbo S ^[2]. Tencent priced the new model at 0.8 yuan per million input tokens and 2 yuan per million output tokens, well below previous Turbo models ^[2], and Decrypt described it as significantly cheaper than the earlier Turbo versions ^[2]. Tencent also said Hunyuan Turbo S doubled word-generation speed and cut first-word delay by 44% compared with previous models ^[2]. First-word delay is the gap between sending a request and receiving the first piece of the answer, which is what users notice most in a chat interface. The new model was reached through a Tencent Cloud API waitlist that included a free one-week trial, and it was not available for public download ^[2]. Decrypt reported an MMLU score of 89.5 for it, slightly above GPT-4o ^[2]; MMLU is a multiple-choice test that spans many school and professional subjects.

Where it sits

The Turbo line shows how competition among model providers moved beyond raw capability. Speed of generation, delay before the first word, and above all the price per million tokens became the figures vendors put in front of business customers, and a cloud platform is the channel through which those customers actually get the model. Naming the successor "Turbo S" and shipping it roughly half a year after the original also shows a quick revision cycle, in which each version is sold mainly on being faster and cheaper than the one before it.