It's got OpenAI on the ropes.

Double Tap

DeepSeek is coming in for the kill.

With Silicon Valley already on its knees, the Chinese startup is releasing yet another open-source AI model — this time an image generator that the company claims is superior to OpenAI's DALL·E3.

Called Janus-Pro 7B, alluding to its beefy seven billion parameters in its full configuration, the AI model was made available on GitHub and Hugging Face to download on Monday, along with a slimmer one billion parameter version. 

It's a followup to an earlier version of Janus released last year, and based on comparisons with its predecessor that DeepSeek shared, appears to be a significant improvement. 

For a quick spin, demos of both its image generation and image understanding capabilities are available online on Hugging Face. Results may vary, but imagery provided by the company shows serviceable images produced by the system.

Bench Warmers

In a technical paper released with the AI model, DeepSeek claims that Janus-Pro significantly outperforms DALL·E3 and another leading image generator model, Stable Diffusion XL, in two key benchmarks: GenEval, in which it boasts a considerable lead, and DPG-Bench, where its margin is much slimmer. Notably, Midjourney was left out of the analysis.

These are only two benchmarks, noteworthy as they may be, and only time and a lot of screwing around will tell just how well these results hold up as more people experiment with the model.

Regardless, DeepSeek sounds adamant that it's onto something big here.

"Janus-Pro surpasses previous unified model and matches or exceeds the performance of task-specific models," the startup wrote on Hugging Face. "The simplicity, high flexibility, and effectiveness of Janus-Pro make it a strong candidate for next-generation unified multimodal models."

In Too Deep

The release of Janus-Pro 7B comes just after DeepSeek sent shockwaves throughout the American tech industry with its R1 chain-of-thought large language model.

Purportedly made on a shoestring budget of under $6 million, DeepSeek's R1 impressively manages to match the capabilities of leading AI models, such as OpenAI's o1, while using just a fraction of the hardware and power.

It wouldn't be an exaggeration to say that the chatbot has embarrassed its foreign competition. Now, serious questions are being raised about the billions of dollars worth of investment, hardware, and energy that tech companies have been demanding thus far.

More on AI: Zuckerberg Convening Huge "War Rooms" to Figure Out How a Chinese Startup Is Annihilating Meta's AI


Share This Article