DeepSeek Piles Pressure on AI Rivals With New Image Model Release

Chinese AI startup DeepSeek has launched Janus Pro, a new family of open-source multimodal models that it claims outperforms OpenAI's DALL-E 3 and Stable Diffusion's offering on key benchmarks. The models, ranging from 1 billion to 7 billion parameters, are available on Hugging Face under an MIT license for commercial use.

The largest model, Janus Pro 7B, surpasses DALL-E 3 and other image generators on GenEval and DPG-Bench tests, despite being limited to 384 x 384 pixel images.

Interesting

[KlomDark]

I played with DeepSeek over the weekend and it worked pretty well on an 8-way Xeon machine with 128 GB RAM and an old 1080ti. Impressive. Looking forward to trying this new image generation model!

Re:

[timeOday]

It's interesting that image models have ended up being so much smaller than language models. No leading LLM could have only 7B parameters or run halfway decently on a 1080ti and 8 CPUs. Intuitively, 'a picture is worth a thousand words' as they say.

Re:

[rwrife]

That was my thought, no one can run the large R1 model at home, but a 7B parameter model can be run by just about anyone.

Re:

[samwichse]

How much video card RAM do you need for smaller a model like that?

Re:

[EvilSS]

8GB should be plenty for a 7B model.

Re:

[samwichse]

Ok, thanks.

Re:

[Rei]

It all depends on what quantization version of the model you use, of course. The lower the quantization, the less VRAM you need, but the more you risk compromising model quality.

Re:

[DogFoodBuss]

You can run the DeepSeek R1 671B at home, as long as you have enough trash laying around with a total of 500GB VRAM. A handful of Mac Minis can do it. https://x.com/alexocheema/stat... [x.com]

Re:

[KlomDark]

The DeepSeek model (text generation not image) I ran was the 8B version. I ran the 70B model, but it was ridiculously slow on my hardware. I'd hate to see how bad the full size version works on it. :) (I'd need a lot more RAM too)

Re:

[KlomDark]

I wrote that terribly. I tried the 8 Billion parameter one and it went acceptably fast, so then tried the 70B parameter one and it was terribly slow. (Maybe a word emitted every 5 to 10 seconds) But at least I got to spin up the fans pretty good!

Re:

[EreIamJH]

Probably a western investment company trying to short AI stocks.

Impressive

[ZipNada]

Maybe this will drive down the training cost for models and the multi-$billion AI shops won't have to build as many nuclear reactors to power their data centers now? It seems like it would be a lot cheaper to figure out how to build these things less expensively than to just dump $500 billion into acres of installations housing Blackwell systems.

Re:

[rwrife]

No, they'll still want nuclear power, they're just going to make the models larger, so instead of the next gen massive model being 1T parameters, I suspect we'll see multi-trillion parameter models by the end of the year.

Re: Impressive

[dj245]

I work in renewable energy. I had some fear in the past couple of years that US renewable PTC and ITC subsidies might be cut, and industry growth severely curtailed. However, big tech has filled any gap that could reasonably come from that. Microsoft, Google, and several others are absolutely dominating the growth of renewables in some parts of the country. If there's a viable project they'll sign a power purchase agreement which is a key element to actually building a large project. Big tech is facilitating and in some cases directly funding a lot of the renewable build out, and those wind and solar farms will still be in place and benefiting the country even if the AI bubble pops and all the server racks go to eBay. The AI boom is kind of dumb, but there are plenty of reasons to appreciate the positive impact this race is having on other sectors.

Re:

[Tablizer]

even if the AI bubble pops and all the server racks go to eBay.

Company Foo now: "We're an AI company."

Company Foo after poppage: "We're a server-farm parts distributor".

Re:

[ZipNada]

>> a power purchase agreement which is a key element to actually building a large project

I think they have to lock in the power arrangements in order to get financing? Most new electricity generation in the US these days is renewable, and the tech industry needs lots of new electricity..

"Solar accounted for 79.3% of all new utility-scale generation placed into service in the first ten months of 2024. In October alone, solar comprised 91.8% of all new capacity added."
https://www.solarpowerworldonl... [solarpower...online.com]

Re: Impressive

[dj245]

I think they have to lock in the power arrangements in order to get financing? Yes that's part of it. Bankers want to see a proposal that has all the revenues on one side of the page, and all the expenses on the other. The other major components of the deal include insurance (my field), a construction contract, a maintenance contract, an approved place to inject the power into the grid, and land agreements for any non-owned land. It is quite an undertaking and the pace of build out is staggering consideri

Re:

[AmiMoJo]

Seems more likely that of all those hundreds of billions that are being invested in AI, a big chunk will get diverted to R&D so they can clone DeepSeek's work. Or maybe the NSA will get involved and try to steal it.

This is good news though. AI used to be only available to massive corporations with deep pockets to invest in training. Thanks to DeepSeek and the fact that they open sourced much of their work, now startups can afford to build their own models. More choice, more variety, and less environment

With really small means ?

[Voice of satan]

Isn't all this just psychological warfare ? To make investors panic. We have no proof that model was trained for as cheap as the Chinese claim. We do not know if they really used a so small number of H800 GPU.

Re:

[rwrife]

Investor panic, this means the limitations we *thought* we had on AI are gone until we hit the next limitation. Imagine using the efficiency gains they used to train on 50k GPU, but scale it up 10-100x and see what kind of AI we'll get.....it's going to be insane. Deepseek didn't hurt nVidia, MS, Meta, Google, OpenAI....they just gave them the method to leap 5 years into the future overnight.

Re:

[Iamthecheese]

Panic? If this is true it's a sign there's more potential in LLMs. Investors will want to get onboard. China released open source versions. It's not going to be all that hard to copy their work.

Re: With really small means ?

[databasecowgirl]

It's also reminiscent of the French response in Monty Python and the Holy Grail, "We've already got one!"

From my limited experience, I must admit the resources needed for chatGPT do seem extreme.

Re:

[oumuamua]

The training rough cost has been verified as accurate, use 6 Flops per parameter token: https://youtu.be/X5adgxV0gBE?s... [youtu.be] Rest of the video goes over tweets from big names in AI

China vs copyright law

[Dr. Tom]

Any details on how the training data were curated?

Re:China vs copyright law

[allo]

From the tech report:
> In Janus-Pro, we incorporate approximately 72 million samples of synthetic aesthetic data, bringing the ratio of real to synthetic data to 1:1 during the unified pretraining stage. The prompts for these synthetic data samples are publicly available, such as those in [43]. Experiments demonstrat that the
model converges faster when trained on synthetic data, and the resulting text-to-image outputs are not only more stable but also exhibit significantly improved aesthetic quality.

Re:

[Mr. Dollar Ton]

Great news, now we can have 10x the disinformation and fakery for the same amount of AI-dollars.

Economic Warfare

[nehumanuscrede]

The only thing China needs to do is to keep throwing out the possibility that their AI is
on par or exceeds that of their Western counterparts and well . . . . you see what it did
to the markets today.

It's crazy how the markets react to the mere mention / possibility of a thing these days :|

Hell, it doesn't even have to be true or verified. Just the suggestion is enough to wreck
havoc.

You Keep Using That Word

[Barny]

open-source

I do not think it means, what you think it means.

Looks unimpressive, IMHO

[Rei]

Most people's whelm seems to be very much under [reddit.com].

" I tried it on huggingface and its...average to good, but I wouldn't say its groundbreaking from what I seen with the few trials I gave it."

"I tried the huggingface stable diffusion(sic) demo and was completely underwhelmed for realistic images. I can only assume config or user error because it can't possibly that bad"

"Does it support nsfw content too?" "If you manage to generate anything that resembles a human, yes"

"Good to see this. But just for text2img pu