Highly acclaimed AI enterprise, DeepSeek, has made waves with its latest release, the Janus Pro family of AI image models. These models, available on the progressive AI development platform, Hugging Face, claim to surpass the performance of OpenAI’s DALL-E 3, and vary in capability with parameters ranging from 1 billion to 7 billion.
Under the open-source MIT licence, Janus Pro proves its versatility by offering commercial use without limitations. Said to be a “novel autoregressive framework” by DeepSeek, Janus Pro exemplifies dexterity as it can dissect and generate new images simultaneously.
Despite only producing images of smaller size resolutions, up to 384 x 384, the Janus Pro clan leaves a bold impression with its compact form and high performance. It not only challenges older models, such as PixArt-alpha, Emu3-Gen, and Stability AI’s Stable Diffusion XL but also surpasses the latter on AI evaluations, GenEval and DPG-Bench.
In a statement made on Hugging Face, DeepSeek proudly acknowledged the Janus Pro’s success in matching or even surpassing task-specific models’ performance as it combines simplicity, adaptability, and efficiency, giving it an edge for becoming an influential model in the future.
DeepSeek, a major player in the AI arena, backed by quant trading firm High-Flyer Capital Management, catapulted to fame this week. Their language models, designed using resource-efficient techniques, raised eyebrows among Wall Street analysts and technologists alike, provoking questions on whether the U.S. can continue its leadership in the AI race and if AI chips demand can be sustained.
Original source: Read the full article on TechCrunch
