UAE Lab Releases Open-Source Model to Rival China's DeepSeek (gizmodo.com) 43
"The United Arab Emirates wants to compete with the U.S. and China in AI," writes Gizmodo, "and a new open source model may be its strongest contender yet.
"An Emirati AI lab called the Institute of Foundation Models (IFM) released K2 Think on Tuesday, a model that researchers say rivals OpenAI's ChatGPT and China's DeepSeek in standard benchmark tests." "With just 32 billion parameters, it outperforms flagship reasoning models that are 20x larger," the lab wrote in a press release on Tuesday. DeepSeek's R1 has 671 billion parameters, though only 37 billion are active. Meta's latest Llama 4 models range from 17 billion to 288 billion active parameters. OpenAI doesn't share parameter information. OpenAI doesn't share parameter information.
Researchers also claim that K2 Think leads "all open-source models in math performance" across several benchmarks. The model is intended to be more focused on math, coding, and scientific research than most other AI chatbots. The Emirati lab's selling point for the model is similar to DeepSeek's strategy that disrupted the AI market earlier this year: optimized efficiency that will have better or the same computing power at a lower cost...
The lab is also aiming to be transparent in everything, "open-sourcing not just models but entire development processes" that provide "researchers with complete materials including training code, datasets, and model checkpoints," IFM said in a press release from May.
The UAE and other Arab countries are investing in AI to try reducing their economic dependence on fossil fuels, the article points out.
"An Emirati AI lab called the Institute of Foundation Models (IFM) released K2 Think on Tuesday, a model that researchers say rivals OpenAI's ChatGPT and China's DeepSeek in standard benchmark tests." "With just 32 billion parameters, it outperforms flagship reasoning models that are 20x larger," the lab wrote in a press release on Tuesday. DeepSeek's R1 has 671 billion parameters, though only 37 billion are active. Meta's latest Llama 4 models range from 17 billion to 288 billion active parameters. OpenAI doesn't share parameter information. OpenAI doesn't share parameter information.
Researchers also claim that K2 Think leads "all open-source models in math performance" across several benchmarks. The model is intended to be more focused on math, coding, and scientific research than most other AI chatbots. The Emirati lab's selling point for the model is similar to DeepSeek's strategy that disrupted the AI market earlier this year: optimized efficiency that will have better or the same computing power at a lower cost...
The lab is also aiming to be transparent in everything, "open-sourcing not just models but entire development processes" that provide "researchers with complete materials including training code, datasets, and model checkpoints," IFM said in a press release from May.
The UAE and other Arab countries are investing in AI to try reducing their economic dependence on fossil fuels, the article points out.
Re: (Score:1)
Everybody seems to hate some group these days.
Re: (Score:2)
Re: Do they still (Score:2)
You might want to learn more about the early days of Islam and what the holly book of Islam says about how to treat people from other religions. That will explain a lot of things.
Personally I think all humans deserve to be treated as humans beings regardless of race, biological sex or religion. Actual behavior is what you should be judged on.
For example I do not approve of genocide, regardless of who is on either side and I'm quite disappointed that my tax dollars are used for genocidal purpose, regardless
Re: (Score:2)
Damn, they've turned Christian without telling anyone.
PhallicGPT (Score:1)
PhallicGPT
Mantra? (Score:2)
Re: (Score:2)
Or information on the murder of their whistleblower.
What is it with that "K2" naming? (Score:2)
Is there something special about "K2" that I am not aware of?
Re: (Score:3)
It's "the second-highest mountain on Earth".
https://en.wikipedia.org/wiki/... [wikipedia.org]
Re: (Score:2)
It's "the second-highest mountain on Earth". https://en.wikipedia.org/wiki/... [wikipedia.org]
Certainly, but why would you opt for such a name knowing others named their LLM also "K2"? They could have been more creative, and named it "Everest" or "LLM Khalifa", or whatever.
Re: (Score:1)
It means Inspector Clouseau's Cato.
This is a good thing (Score:2)
Having one country or monopolist own the tech is bad
The best course of action is for the tech to be shared by all, preferably open source and open weight
Monopoly is inevitable (Score:2)
And two things are going to destroy that training data.
First sights are gradually walking down to block AI bots if only because the excess traffic is crushing them.
And second the internet is filling up with AI slop and if you train your AI on slop you're going to get limitless slop
This means that before long the only people who will be able to maintain useful llms outside of a handful of extremely specific scientific purp
Re: (Score:2)
Why have one authoritarian country have all the AI when 3 authoritarian countries can have all the AI.
(Sorry americans, but right now I'm counting you guys as an authoritarian country. With that said, if you guys arent using that constitution anymore, can we have it? )
Re: (Score:1)
You've been drinking authoritarian koolaid. We've just started using our Constitution again here in the states and we tossed the authoritarians. We seem to be making some inroads in helping the UK and Canada ditch their authoritarians as well.
Flawed Reasoning (Score:5, Interesting)
I notice they cite "math performance" as being a bright spot. Presumably they aren't talking about speed, but correctness.
Let that sink in: We now have to question not only how quickly our computers can perform a calculation, but whether the calculation is performed correctly at all.
I also notice the term "reasoning model" being applied. Everything I've read indicates the "reasoning" is just a differently-weighted LLM layered on top of the existing one. Artifice on top of artifice, with no real logical anchor. It sounds like exactly the right way to wave your hand for maximum voodoo, though.
Re: (Score:2)
Let that sink in: We now have to question not only how quickly our computers can perform a calculation, but whether the calculation is performed correctly at all.
Yeah, just like with those "quantum computers". Maybe by running some LLM on a "quantum computer" we can have incorrect incorrect results... so, maybe correct ones?
Re: (Score:3)
Schroedinger's techbro. Until his answer is actually tested, it exists in a superposition of states. It is both right and wrong, charming and strange, confident and worthless.
Re: (Score:2)
Your argument is misleading. Math is about theorems, not about calculations.
Before LLMs your computer did not prove theorems at all. Now it can prove a lot of them correctly.
So if that was a good faith argument: LLMs are not made for calculations and should not be used for them (even though recent ones good okay with them), but they are made for things like proving theorems and they are quite fine. In general LLMs are better in explaining math to you than in calculating examples for you, because they are ma
Re: (Score:1)
"Before LLMs your computer did not prove theorems at all. Now it can prove a lot of them correctly."
Actually that was a bit of hype stemming from one success. The hallucinations and high error rates we see in other types of content have been show to still be present in gpt5 proofs.
The thing still doesn't think, reason, or walk through anything... it still just mindlessly spits out a search engine result listing using a statistics based algorithm. The only reason it looks like reasoned output is because the
Re: (Score:2)
I'd say the whole "PhD level" thing is a bit of marketing, but if you consider what the models actually can do, it's quite impressive. I mean, don't you think saying "The proof still contains a mistake" is, on the other hand, acknowledging that the thing that initially was just modeling language managed to write a nearly accurate mathematical proof? Somehow, people seem now unable to be amazed by anything less than perfection, even though most of them are not able to write the proof themselves.
Re: (Score:2)
There is no "almost" or "close" in math, it's either correct or not. You can look at the wall of text it tells you is a proof, be amazed, and then what? If you actually want to do anything with it, you have to find a mathematician to verify it. I don't imagine they're so short of ideas that an "incorrect proof generator" would help them in any capacity. Their own brains come up with incorrect proofs just fine. The entire field of mathematics is in pursuit of separating the correct from the incorrect. It's k
Re: (Score:2)
But there is an almost in people or LLM getting things right. That was the point. You've got a machine that does things most people can't and then it is criticized because it is not perfect in all disciplines.
"They see something that might look human from an angle, and treat it then as a human"
That's the point. Just because the interface is human language and it is styled in a chat format for easy access you shouldn't judge it like a human, but like a program. And it is amazing for a program. Try to build s
Re: (Score:2)
My criticism wasn't so much that it's "not perfect", but that it's functionally useless.
Still not sure what I'm supposed to do with an AI-generated simulacrum of what sounds like a mathematical proof, other than gape at it. Not sure what a mathematician would do with it either.
Re: (Score:2)
I don't see it inventing bigger new proofs yet, but I see it being useful to brainstorm things. Most proofs rely on a lot of well-known properties and the models know many of these. So a short question "Does X hold" can help a lot. You either verify it yourself or look it up, but you get first feedback to bounce ideas. And I think the reasoning models are particularly useful, because in their reasoning trace are a lot "Maybe this, no rather that, wait let's look at in from another perspective" that can give
Re: (Score:2)
Maybe the thing about the almost was not that clear.
Let's do an example: You hold a math tutorial and check the proof of a student. You see they got six of seven steps in the proof right. Then it is almost correct.
The problem you have with the LLM is that you expect perfection. Instead of saying "It is the level of a fourth semester student" you say "It is useless, because the seventh step of the proof is wrong!" Nobody said it will get every proof correct. If you look at the list of retracted papers, you'l
Where's the link? (Score:2)
https://chat.deepseek.com/ [deepseek.com] is for deepseek
Where's the link for K2?
I searched and it seems like it's only available for pay by other websites?
Re: (Score:1)
Re: (Score:1)
https://www.k2think.ai/ [k2think.ai]
Missed opportunity... (Score:3)
...to call it DeepSheikh.
Re: (Score:2)
I wish I still had modpoints for you.
How is this compatible with Islam? (Score:2)
I mean, playing God is shirk, the greatest sin and crime against God in Islam.
Re: (Score:2)
LLM is God... Said nobody ever, even Sam Altman. It's creating intelligence (and life) that some people consider "playing God", and the industry goal of creating AGI arguably comes close to that.
I personally don't think current AIs qualify as either life or intelligence.
Debunking the Claims of K2-Think (Score:2)
Maybe interesting for some as additional information: https://www.sri.inf.ethz.ch/bl... [inf.ethz.ch]
In general: If something sounds too good to be true it often is.
OpenAI doesn't share (Score:2)
> OpenAI doesn't share parameter information. OpenAI doesn't share parameter information.
I am glad the summary accidentally wrote this twice. OpenAI was founded as a non-profit with a goal of being OPEN and has since become neither open nor non-profit. I refuse to use anything OpenAI for this reason and I wish others would reject them as well. Integrity matters. Demand Open.