DeepSeek's New Image Model: A Victory for Affordable AI?

DeepSeek's Momentum Surges with Janus-Pro Release

Chinese AI startup DeepSeek is certainly not letting its momentum wane. Hot on the heels of dethroning ChatGPT from the top spot in the App Store for most downloaded free apps, the company unveiled its latest innovation, the Janus-Pro multimodal text-to-image AI model, on Monday. Much like its flagship model, R1, Janus-Pro is available under an open-source MIT license, making it not only freely accessible but also commercially viable. Enthusiasts and developers can download it from HuggingFace and GitHub.

Also: I've put DeepSeek's R1 and V3 coding skills to the test, and while there's room for improvement, it's clear we're not facing AI apocalypse just yet.

Following the footsteps of R1, DeepSeek didn't stop at just one version of Janus-Pro. They've rolled out several iterations, ranging from 1B to 7B parameters. According to DeepSeek's internal testing, Janus-Pro-7B, the heftier model, outperforms well-known image generators like Stable Diffusion and DALL-E on the GenEval and DPG-Bench benchmarks. The company touts Janus-Pro's use of an "autoregressive framework" which they claim "surpasses" unified models.

Building on Success: Janus-Pro's Capabilities

Janus-Pro evolves from its predecessor, Janus, which was released last year. This new model not only creates but also analyzes images. However, there's a catch with the smaller-parameter models; they're limited to analyzing images at a resolution of 384 x 384, which could be a limiting factor for some users.

Despite this, Janus-Pro holds its ground impressively, especially considering DeepSeek's reportedly lower training costs compared to those of its US counterparts. A research paper from December stated that their V3 model was developed for just $5.6 million, a figure that pales in comparison to the budgets of Google and OpenAI's leading models. Skeptics, however, question the completeness of this cost estimate, pointing out potential exclusions like R&D, data, and personnel expenses.

Industry Impact and Future Prospects

Even Nvidia has chimed in, with a spokesperson telling CNBC that Janus-Pro represents "an excellent AI advancement." Given DeepSeek's rapid release cycle, the initial feedback on Janus-Pro has been mixed but largely positive. As more users get their hands on it and compare it with other image models, these impressions may evolve.

Also: The secret behind DeepSeek AI's success has been uncovered by Apple researchers, shedding light on their innovative approach.

ZDNET is currently investigating claims that DeepSeek's methods are more energy-efficient than those of its US competitors. If true, this could be a game-changer for the AI industry and how investments are directed. The launch of Janus-Pro also raises questions about the necessity of grandiose projects like Stargate, a $500 billion collaboration between multiple AI giants, which was promoted by the Trump administration. With competitive AI models like Janus-Pro, the need for the energy-intensive and massive scale of Stargate's proposed data centers comes into question.