Tencent has just launched a new model of AI capable of Generate images from a text prompt. Traditionally the proprietary models have dominated this type of creative task, but the Tencent model has given the surprise and according to various benchmarks is able to generate images better than the rest of competitors, including those of Google and Openai.
Hunyuan Image 3.0. This is the name of the new Tencent model, which in the LMARENA classification Of more powerful models for the generation of images from text has managed to overcome Gemini 2.5 Flash image preview (popularly known as Nano Banana), in addition to other proprietary models such as GPT-IMAGE-1, FLUX-1-KONTEXT-MAX or QWEN-IMAGE.
The Tencent image model has already managed to overcome its competitors according to LM Arena experts.
Blind vote. This LM Arena classification works through a blind voting system in which the users choose their favorite images without knowing what model generated them. And according to this vote, this Tencent model surpassed all its opponents, including Google’s popular “Nano Banana”. Of course: the vote also takes into account long -term results, and the short period that has been available Hunyuan Image 3.0 makes the results qualify as “preliminary” and not definitive.
How it works. Those responsible for Tencent explain in the description of the model How they have used a new diffusion architecture that makes use of dual encoders (a multimodal llm and another that better understands characters in different languages) and RLHF optimization (Reinforcement Learning from Human Feedback, which refines the previous result) for the creation of higher quality images. The system makes use in addition to a compression system so that the entire process consumes less resources without loss of quality.
Open pesos and commercial license. Hunyuan Image 3.0 is a model that shares Your code in github and that offers A license surprisingly permissive. In fact it is possible to use it for commercial and professional purposes.
The price is not entirely economic. Although the model can be tested for free in the Project websitein our tests we could only create an image (10 credits). The platform allows you to buy monthly credits: $ 8 per month allow Buy 500 creditswhich a priori would allow us to create 50 images of 10 credits each. Each would go to $ 0.16, when Nano Banana It has a cost of $ 0.039, four times lower. There are other options to try it, such as Hugging Face “Spaces”. It is also possible to get an API key In Tencent Cloud To use the locally.
Gemini continues to win as “editor”. Although the Tencent model is interesting and remarkable, Nano Banana continues to win the game if we consider that it has become a unique substitute for the traditional Photoshop. Many users no longer edit photos but carry one in Gemini and then they tell the AI what changes they want to make in that image.
Alibaba, more conversational. Although Hunyuan Image 3.0 can allow something like that – in fact There are demos In this regard – the interface is now more aimed at a single prompt to generate images, not to a “conversation” as GEMINI allows. Another of the protagonists of this land is Alibaba, who with Qwen-Image-Editor adopts the same approach as Google with Gemini and Nano-Banana. In that Alibaba model “you speak” with your image to ask for changes, something that at the moment does not seem that the Tencent model does so directly (although it does not seem difficult for him to get it).
But be careful. The differential here is that the generation of images, which seemed to be dominated by proprietary models, can be apparently equally good (or even superior) through open models. One more time The Chinese commitment to that philosophy is remarkable and contrasts with the closed approach and owner of most American companies that develop AI models to generate images and text (or, of course, video).
Image | Hunyuan
GIPHY App Key not set. Please check settings