Google battles Chinese open-weights models with Gemma 4

Google on Thursday unleashed a wave of new open-weights Gemma models optimized for agentic AI and coding, under a more permissive Apache 2.0 license aimed at winning over enterprises. The launch comes amidst an onslaught of open-weights Chinese large language models (LLMs) from Moonshot AI, Alibaba, and Z.AI, many of which now …

[Paul Herber]

Does anyone fancy trying to use one of the Chinese models in Tibetan?

[cyberdemon]

Does anyone fancy using one of the American models in Persian?

It's a serious point though: So-called Frontier Models are being used as if they are arbiters of truth and logic, when they portray distorted truth, and hollow imitation logic. The only explanation I can conceive for nations to incinerate trillions of dollars of resources on this toxic bonfire, is because it promises to replace truth and logic with a simulacrum which they can influence, and provides total surveillance of the innermost thoughts of those who have forgotten how to think for themselves.

If everyone outsources their thinking and research to an LLM, then who needs a mind-control-ray to keep the populace at bay?

[_wojtek]

e4b

Tried locally gemma4:e4b on my oldish MBP M1 and it was surprisingly decent...

I wouldn't be surprised if, INA couple if years, majority if such AI would be run locally given the advances... (and hopefully the market would get over current shortages of ram by cranking up production...)

[Yorick Hunt]

Re: e4b

"... I wouldn't be surprised if, INA couple if years, majority if such AI would be run locally... "

I'm sorry Dave, I can't let you do that. There's no way the "tech bros" are going to let their cash cow troddle off to greener pastures.

[_wojtek]

Re: e4b

well, they are already doing that (obviously not out of the goodness of their hearts)

[werdsmith]

Re: e4b

Run locally but with processes for local training so optimised for the particular use.

So rather than trying to answer every question , they will use all resources dedicated to the stuff the user actually needs.

[Anonymous Coward]

The image processing is insane. I cannot believe I'm getting these results in only 4GB VRAM (for the 4B model). It can tell me the intent and reasoning behind an image, make up stories about them, it's like it can really see the image. It's definitely not fool-proof of course, and there are some subtle things it doesn't grasp even when told specifically what to look for, but I'm blown away it's doing all this on only 4GB. I've been following open LLMs since they became a thing, and this is the first time I've been genuinely impressed ever since Llama 3 came out. I'm really excited to see what else it can do.

I haven't messed around with actually talking to it yet, but it seems very akin to regular Gemini, likes to talk in verbose lists more than anything.

[IGotOut]

"It can tell me the intent and reasoning behind an image, make up stories about them, it's like it can really see the image"

I can taste the vomit in my mouth and I've not even had breakfast yet.

Meanwhile, tomorrow I'll be interacting with a bunch of real human artists, in a pub, talking about real human issues, and how real humans come up with original art, based on human interactions and experiences. Then I'll be out partying until dawn, interacting with real human beings, creating real memories that will make me smile years later. Your image? Probably forgotten in a few days or weeks.

Feel free to generate your slop and think it's the same thing, because I know it's not.

[werdsmith]

Will all the other humans be as sanctimonious as you?

[Anonymous Coward]

I'm just playing with a chatbot, jesus christ.

[TheMaskedMan]

"I'm just playing with a chatbot, jesus christ."

Hmm, now there's an idea: JC resurrected in LLM form, just in time for egg day. Could be a winner!

[TheMaskedMan]

"Feel free to generate your slop and think it's the same thing, because I know it's not."

And you were so eager to tell us all about your social life, such as it is, that you failed to read what the OP actually says. Nowhere did they say they were generating images , just analysing them. Perhaps a little less froth-at-mouth and a little more think-with-brain might help - if you can't manage that, I'm sure there's an LLM to help you.

I, however, do generate images, sometimes in copious quantities. And then, if I feel like it, I sell them in various online bazaars without ever having the misfortune to interact with temperamental arty-farty wetware. I don't give a damn if they're forgotten, or even never looked at - it's a pretty picture, something to fill a space on a wall or website, nothing more.

[blu3b3rry]

I'm amazed the AC managed to type all that and keep a straight face.

I'm assuming the above is satire, anyway.

[cd]

Good to see MoE still working, and cURLy...don't mention Larry. I don't like Larries.