Viral AI company Deepseek has released a new set of a multimodal AI model that claims to be better than Openai’s Dall-E 3.
The model that can be downloaded from the face of the AI dev platform is part of the new model family called DeepSeek called JANUS-PRO. The size is 1 billion to 7 billion parameters. Parameters are almost compatible with model problem solving skills, and models with more parameters generally have a better performance than those with fewer parameters.
JANUS-PRO is under the MIT license. In other words, it means that it can be used commercially without restrictions.
Deepseek describes “novel self-regression framework” Janus-Pro can analyze and create new images. According to the company, the two AI evaluation benchmarks break the Geneval and DPG-Bench, the largest JANUS-PRO-7B, and breaks Dall-E 3 and breaks Pixart-Alpha, EMU3-GEN, and Models. Stable diffusion XL of stability AI.
Certainly, some of these models are on the old side, and most JANUS-PRO models can analyze only small images at a resolution of up to 384 x 384. However, the performance of JANUS-PRO is impressive considering the compact size of the model.
“The Janus-PRO exceeds the previous unified model and aligns or exceeds the performance of a task-specific model,” says Deepseek’s posting on his face. “Due to the simplicity, high flexibility, and effectiveness of Janus-Pro, it will be a powerful candidate for the next generation unified multi-modal model.”
Deepseek, a Chinese AI lab, which was mainly funded by quantitative high -flight capital management, has entered the mainstream consciousness this week after the chat -bot app rose to the top of the Apple App Store chart. 。 Deepseek’s language model is trained using good calculation -efficient technologies, and many wallstre trex analysts and technicians have maintained whether the United States can maintain a lead in AI races, and the demand for AI chips. I came to question whether or not.
Update: The earlier version of this story implies that the janus-pro model can only output only small (384 x 384) images. It is not true. I regret the error.
TechCrunch has a newsletter focusing on AI! Sign up here and get it on the reception tray every Wednesday.