Didn't know the "Highest scoring EU model" has a low bar, lower than Qwen3.8-27B, but still congratulations on the milestone, hopefully next iterations will get better from here
It is surprising given how many parameters it has that it scores so low. But, hopefully this will build up domestic talent and understanding and let Europe compete on the world stage with this.
3.8-flash-next quantized in a "large" Q4 that just fits in 128GB RAM even more so, in how close it can get to state of the art in a number of benchmarks. Or a large Q8 version of it that fits in under 190GB. Competing against things that are closed weights/opaque information about the model and might very well be 600B+ in size.
> the state that has to buy all its water and food from its aggressive, militarized neighbour.
As a complete tangent, now imagine being a Canadian and realizing how much of your fresh fruits and vegetables come from the USA (or if from Mexico, through the USA).
That's a very limiting view of things. A model, even an open weight one, is never neutral, it is an encoding of a way of viewing the world.
What kind of "alignment" are AI labs optimizing for? Ideological alignment is the full term, self-censored into something more technological-sounding.
Every model has people behind it rating what it should and shouldn't say. Every time you ask a model and trust its answer, you become ever-so-slightly ideologically indoctrinated.
I don't want my model to reflect the views of American oligarchs or Chinese cadres. I want European values of enlightenment and humanitarianism to be the default.
>I don't want my model to reflect the views of American oligarchs or Chinese cadres. I want European values of enlightenment and humanitarianism to be the default.
I don't want AI models to reflect any kind of values whatsoever. I have my own views - thank you - and I don't need other throwing their values in my face using AI.
I prefer AI models not being trained ideologically.
Unfortunately, AI models are trained on humans, and so they're getting that whether we like it or not. I'd also prefer it to be neutral, but I don't think it's possible and the second-best option is to have models from multiple regions and balance things out.
The problem is always the next model, or the one after. If china thinks it's beneficial to stop open weight releases, it will stop them. Then you are stranded on that one and no local industry to produce new models for you.
I hope you're aware that it's likley in the future open weight models will output tokens/commands that aren't in your interest. It could be that models are trained with spionage in mind and models could only target specific input token patterns, time zones, ip addresses, names etc... You will be able to see the output tokens but things will become so complex that you won't notice its intentions. Maybe there will be some sort of output token scanning software/llm (let's call it a modern form of a virus scanner) that validates intentions. In the end, I'm just trying to say that open weight means you have a black box in front of you that you don't know what it does. Therefore, the training material of open weight models should be known. While it doesn't matter where the model comes from, it certainly matters whether you trust those people if you don't have the training material.
It's more important to have ownership into them, they're turning into powerfull tools that we have already seen cut off in the private sector on a whim. Having our own that is comparable (this is a nice first step) is better than Alibaba for example just not releasing the rest.
As with all of the EU's dependancies on Tech with the US, the idea of "I don't care where my DNS servers are" is starting to be an oversight.
As a non-European, I care. I definitely want choice in this matter, as too many things can be baked into models that we can't really know until it comes up.
For instance, I heard one of the Chinese models has some interesting "history" "facts" built into it. That's major, and can influence a lot more than just asking it for that particular history lesson.
I don't want just 1 LLM. I want a great LLM from every major region. I will probably still prefer my own region, but I still want the others to be an option.
If these things replace search engines, you definitely don't want that technology to originate from China or other country where "truth" comes from Dear Leader
not being technically independent means you'll end up being a serf when American or Chinese companies decide that you don't have access anymore for geopolitical reasons
This is an oft-expressed refrain. But there are good reasons to not want to use Chinese models (or models from any adversary for that matter).
1. They may be trained, in theory, to inject subtle back-doors into certain kinds of generated code.
2. They may be also trained to include back-doors when deployed in public-facing services where user can provide text or image input.
3. Chained with (2) they may be also trained to exploit their inference environments, which though a big feat, not outside the capability of nation-state hackers.
What I want is a model that is trained with data that is openly available, where the data is curated by academia. I don't want corporate crap in my AI (unless it has been filtered properly).
I do not want open weight models from China to be the only viable locally hosted things (deepseek v4 flash 0731 Q8, qwen 3.8-flash-next Q8, GLM-5.3-Flash) in the under 200GB RAM class.
I want to see things like Mistral and Laguna (non-CN) succeed. I have spent about a week using Laguna S 2.1 as a test and while I wasn't blown away by its capabilities, it's also totally acceptable for many purposes.
I do hope these Quasar people learn that if you announce a new model and it already performs worse than things people can go download from huggingface, and/or buy access to with very cheap token plans via openrouter or opencode.. If your new model is API only and people can't download/examine it, it will get very little uptake and real world use.
I can see it as a niche market for european sovereignty stuff if absolutely necessary, hosted and run in Europe, sure. Same as Mistral. That's a niche which exists, there's probably enough room for a couple of modestly sized companies doing it... I guess?
I understood from informal chatter that researchers in the top Chinese labs are pretty open with sharing knowledge with each other. Additionally, it seems that anywhere between 30-50% of key researchers in the top US labs are ethnically Chinese. I wonder if this situation might give Chinese labs/researches some advantage just due to language and informal networks. Chinese researchers can understand all the English research, but research in Chinese is far less accessible to non-Chinese.
Chinese ML researchers primarily publish in English and only secondarily in Chinese. For example, take the Qwen-3.8-Next blog post https://qwen.ai/blog?id=qwen3.8-flash-next (which apparently doesn't include the language choice in the URL, so you'll need to switch to the 简体中文 translation manually). Even in the Chinese version, the "Hugging Face", "Tech Report" and "FlashQLA" links point to English documents, and the ModelScope link has a brief flash of English content before autotranslation kicks in to turn it into Chinese. I'm not sure what is used on the Qwen Discord, but I would guess it's a mix of languages.
Personal communication is of course different from official documentation, but a researcher who wants to establish a working relationship with Chinese colleagues could easily do so while communicating entirely in English.
It's not a small niche. Anything touching European resident personal data must only be done by Euro AI Act compliant AIs. So, hosted in Europe at least (unclear to me rn, would love to learn the exact criteria).
I wish for this company to have great models. I am glad to see such good scores.
I see they have a HuggingFace account[0] and they fine-tuned GPT-OSS, Nemotron, and Qwen, in the past under new names.
There are some things about it that make me worry though.
• They don't indicate the active parameter count, or indicate whether they pretrained the model. It could be a MiniMax M3 finetuning, as the parameter count almost matches (435B vs. 438B).
It would not be the first company with a splashy release, like Brampton Intelligence[1], or SubQ[2].
Unlike those, they do seem to have experience fine-tuning models. Regardless of my worries, I am rooting for them to learn how to train models.
"quantum algorithms" that would put them directly on a blacklist if I had one. They might as well directly promote their stuff with Cold Fusion and Snake Oil.
As a European, or in general, this make me happy. Since the more diversity the better.
Tho it’s hard not to not to think of this as a big fish in a small pond situation (when talking about best model in EU).
When the weights are closed I don't believe any benchmark.
I just got Qwen3.8-27B to score extra 10% on SWE Pro by adding a proxy in front of it that has few simple "harness like features":
- when the model gets stuck it tells it to "go on"
- when it sends no output, malformed json, slips to wrong tool use format, etc it asks it to "try again better"
- detects repetition and tells the model.
- injects a prompt about "planning tool use" when it seems to be using same tools repeatedly.
- injects a reminder it can use tools if there are no tool uses for over X messages.
10% - with just that.
I have more to test. My point is, open weights models get tested on naked model quality. "Frontier" models get tested as a model + whatever secret sauce they choose to put in front.
The company sounds like a bunch of hot air to me. From their about page:
> At the heart of Multiverse's platform is CompactifAI, a compression technology that applies tensor networks, a mathematical framework from quantum physics, to the problem of AI model compression. This application was pioneered by co-founder and Chief Scientific Officer Dr. Román Orús and reduces the size of large language models by up to 80-95% with immaterial accuracy loss.
Ironic, considering they are releasing a 438B model that loses to a 27B one. From another part:
> Singularity Machine Learning is a cloud service that uses quantum machine learning for solving supervised learning problems.
I wouldn't be surprised if these guys just finetuned an open Chinese model and called it a day.
They had other models a while ago, named 'Pulsar', and they were finetunes of Nemotron in collaboration with NVIDIA. I think this will be something similar.
> I wouldn't be surprised if these guys just finetuned an open Chinese model and called it a day.
Easy enough to find out, ask it a whole bunch of questions about politically sensitive things that would be impossible to publish on CCTV, the Peoples Daily, CGTN, etc. If they didn't train the model and just fine tuned it, a lot of "don't talk about Tibet or the Dalai Lama or what happened in 1989" will be perma baked into it.
I instinctively distrust any company that makes their national origin the main selling point. To me that's a subtle signal that they cannot compete on technical merit.
I mean, yes, but some governments may be tied to using European alternatives by regulations and some individuals may prefer not relying on US / China models for various reasons (could be political / personal preferences).
This is a universal problem, interesting enough especially you'll notice that in the automotive sectors a lot. Crappy local automakers pull the nationality card because they can't compete on real engineering.
There's been a whole thing recently with companies boasting about ensuring American supremacy. I assume its a play for Pentagon dollars, but it is the same kind of thing really.
Same with "the original." If their only claim for why I should choose them is they got there first, maybe they should've spent that head start on being better. That said, in some industries it's advantageous to work domestic even if the product is worse. Just a risk vs reward thing
Or, it's a signal that you're not giving your data/money to an untrusted ally.
Do you even understand just how much Europe distrusts the US with its current administration at this point? Nobody wants to deal with US companies if they can.
On the flipside imagine if Europe could be a genuinely competitive continent again. I don’t expect it to happen in my lifetime - but if it could, the benefits to the consumer would be enormous.
Just because you haven’t heard of them and they’re not American doesn’t mean they’re hot air.
I got the privilege to meet their CEO in Paris as they were announcing their partnership with Axelera and we exchanged some stories. They are full of very smart people, and they got accepted in the EIC Accelerator which, while you may not have heard of it, is one of the most rigorous and difficult programs to get into as a startup (far more difficult than YC which nowadays accepts any vague AI trash and always shoots wide).
This is the classic EU thing - the much superior most rigorous program - that no one has ever heard of, that turns out products that are rebadges or clones of other stuff. China has been getting more and more impressive. These EU "experts" are literally rebadging Chinese open models. 15 years ago that would have been a wild concept. China is totally OK fast following with no shame. SpaceX clone, they are getting to re-usability faster than the EU etc.
Except that no, it doesn't churn "clones"; MVC is quite a serious company and working on far more interesting stuff than whatever shit the SaaS or ads company you're statistically likely working at is.
EIC churns out DeepTech, invests 2.5M nondilutive with funding horizons of 10+ years and larger equity-based programmes too. Multiverse is a quantum computing company working on energy-efficient AI algorithms. The EIC is the reason companies such as ICEYE (https://www.iceye.com/) can exist. The chips running all that inference going in those US datacenters being built? 99% certainly european chips, almost certainly EIC-funded. (Eg. Hailo, Axelera, ..)
You haven't heard about the EIC because they don't care about marketing to americans; their target market is European companies, not you. But sorry to say, the talent is in Europe; the only reason you've been getting that talent in the US is a better business ecosystem, and your current administration has been utterly destroying this while our leaders have been actually fixing our issues. So keep shitting on Europe from the top of that crumbling hill, we'll see you when you emigrate.
Which states that this is a 40 TOPS int4 PCIe device, not what you'd put in a datacenter and certainly not what can run a frontier class model (or the model above). Embedded devices for inference are really cool! But that's the opposite end of scale for what goes into a data center.
I was sharing the EIC backed ones, but it's fair that in a datacenter it's more likely to be Axelera (or vsora/semidynamics if you start digging into earlier-stage competitors). Nvidia is the only American company selling AI chips to datacenters at scale but they're a huge supply chain risk so it's most DCs don't do single-supplier either way.
You'd need to compare by volume. And when you do that - unfortunately - NV is by far the largest supplier. 'Most DCs' are effectively single-supplier. And I wish it weren't so. This is the hard one to crack, NV needs credible competition. AMD and Intel are the only two majors that potentially could do this but they're both way behind where they should have been today.
I remember doing my first bit of 'shader repurposing' a long time ago thinking 'wow, I wonder whether Intel realizes this graphics card is the more powerful processor in my computer'. Back then (in the pre-CUDA) days you had to contort yourself in the worst possible ways to get any kind of unintended computation done but when it worked it was amazing. Since those days NV has been pumping enormous effort into building out that miniscule head start that they had in the days of 'Brook' and the subsequent release of CUDA. It is going to cost a lot of money to duplicate that, and even more to compete with NV on what is now their home turf. But I really wished someone with very deep pockets would do just that.
I might be misreading this. But they seem to heavily imply that they trained this model, start to end. Or at least want to give off the impression that that is the case.
A company, whose bread and butter is to remove parameters from models [1], releases a 438B model. Their wording is a little slippery, but they don't actually say they trained it.
I think this is GLM 5.2 with parameters removed. Its advertised in their changelog [2] as
"capabilities are identical to GLM 5.2," it has the same two effort settings "high" and "max," and both are text only [3]. I might be wrong about this, but I would love to hear more about what they did before changing my mind.
Not the greatest fan of the marketing personally. Irrespective of what this model is.
On Artificial Analysis it's listed as "Quasar 438B (max, based on GLM-5.2)" - so you see exactly right. Not sure if this was changed post publicity drive or not, this is the first I'm seeing about this model.
I don't remember that being there, but you know how memory can be. I checked the Wayback Machine and it has a snapshot of the page that does not mention it. The snapshot is from September 1st, not from yesterday, so perhaps they changed it before this announcement and my comment was made. But I personally don't think so.
Multiverse Computing is one of the weirdest companies I've encountered here in Europe/Spain. Their product / job application descriptions are just technobabble.
Despite my doubts I applied to one of their positions a couple of years back only to receive a super late and generic "we are not moving forward" mail (I'd say I fit pretty damn well for the position, but it seems to be the new normal that most companies don't even want to chat with people 99% of the time).
I don't get what the issue is. Chinese labs fully publish how they make great models. Architecture is known, training methods are often very open. Why Europe just can't copy what they do?
Their CMO's bio lists "20,000 'qualified' quantum AI contacts on linkedin" as the SECOND line in his bio.
"Business Management by ESADE. Co-Founder and CMO of Multiverse Computing. President, “barcelonaqbit-bqb”, 20,000 “qualified” quantum AI contacts on LinkedIn. VP of the AMETIC Innovation..."
Isn't this standard in Europe? There exists a layer of well-connected and impenetrable managerial class, shielded by a glass ceiling. The entire company is BS, and they use some folks with clout and credentials to access grants and handouts (or whatever inscrutable financial reasons). They often sell products to each others' companies or big Euro vendors.
Even if this fits the spirit of corruption, I'm not sure it fits the spirit of it. They certainly aren't there to build a product or provide valuable expertise, but taxes and salaries are paid, people go to conferences, they might even build something, but its a giant LARP. Don't expect to get rewarded or recognized for solid engineering.
I'm sure if you worked for a mid-sized Euro firm, you know what I'm talking about.
So "Europe's Leading AI Model" is just a modified Chinese model which does worse than the better Chinese models?
It seems Europe is many years behind in tech. Maybe for AI models they are just a decade behind, but as far as producing hardware capable to run SOA models they lag tens of years.
It is beyond me how they managed to acquire such funding. Sounds a lot like an earlier quantum computing effort which pivoted to "AI" as well: https://zapataquantum.com
92 comments
[ 0.19 ms ] story [ 18.9 ms ] threadAs a complete tangent, now imagine being a Canadian and realizing how much of your fresh fruits and vegetables come from the USA (or if from Mexico, through the USA).
What kind of "alignment" are AI labs optimizing for? Ideological alignment is the full term, self-censored into something more technological-sounding.
Every model has people behind it rating what it should and shouldn't say. Every time you ask a model and trust its answer, you become ever-so-slightly ideologically indoctrinated.
I don't want my model to reflect the views of American oligarchs or Chinese cadres. I want European values of enlightenment and humanitarianism to be the default.
I don't want AI models to reflect any kind of values whatsoever. I have my own views - thank you - and I don't need other throwing their values in my face using AI.
I prefer AI models not being trained ideologically.
E.g. I had random, completely unrelated and irrelevant fetch by Qwen3.8 27B to " https://routify-file-proxy-sg.oss-ap-southeast-1.aliyuncs.co..." and I only noticed it because I have allowlist rules for what they can do.
As with all of the EU's dependancies on Tech with the US, the idea of "I don't care where my DNS servers are" is starting to be an oversight.
But thanks for your opinion.
For instance, I heard one of the Chinese models has some interesting "history" "facts" built into it. That's major, and can influence a lot more than just asking it for that particular history lesson.
I don't want just 1 LLM. I want a great LLM from every major region. I will probably still prefer my own region, but I still want the others to be an option.
Having some knowledge about the training data is useful.
not being technically independent means you'll end up being a serf when American or Chinese companies decide that you don't have access anymore for geopolitical reasons
if sovereignty is the goal, drop Microsoft, aws and adapt / develop open source / EU alternative.
once serious progress is made there, more resources can be redirected into AI after all it's already too late to join the race.
1. They may be trained, in theory, to inject subtle back-doors into certain kinds of generated code.
2. They may be also trained to include back-doors when deployed in public-facing services where user can provide text or image input.
3. Chained with (2) they may be also trained to exploit their inference environments, which though a big feat, not outside the capability of nation-state hackers.
I want to see things like Mistral and Laguna (non-CN) succeed. I have spent about a week using Laguna S 2.1 as a test and while I wasn't blown away by its capabilities, it's also totally acceptable for many purposes.
I do hope these Quasar people learn that if you announce a new model and it already performs worse than things people can go download from huggingface, and/or buy access to with very cheap token plans via openrouter or opencode.. If your new model is API only and people can't download/examine it, it will get very little uptake and real world use.
I can see it as a niche market for european sovereignty stuff if absolutely necessary, hosted and run in Europe, sure. Same as Mistral. That's a niche which exists, there's probably enough room for a couple of modestly sized companies doing it... I guess?
Personal communication is of course different from official documentation, but a researcher who wants to establish a working relationship with Chinese colleagues could easily do so while communicating entirely in English.
It's not a small niche. Anything touching European resident personal data must only be done by Euro AI Act compliant AIs. So, hosted in Europe at least (unclear to me rn, would love to learn the exact criteria).
I see they have a HuggingFace account[0] and they fine-tuned GPT-OSS, Nemotron, and Qwen, in the past under new names.
There are some things about it that make me worry though.
• They don't indicate the active parameter count, or indicate whether they pretrained the model. It could be a MiniMax M3 finetuning, as the parameter count almost matches (435B vs. 438B).
• They mention using quantum algorithms in other projects: https://multiversecomputing.com/singularity despite quantum algorithms not being typically useful currently.
It would not be the first company with a splashy release, like Brampton Intelligence[1], or SubQ[2]. Unlike those, they do seem to have experience fine-tuning models. Regardless of my worries, I am rooting for them to learn how to train models.
[0]: https://huggingface.co/MultiverseComputingCAI
[1]: https://x.com/newsystems_/status/1904577550690771050
[2]: https://subq.ai/introducing-subq
I just got Qwen3.8-27B to score extra 10% on SWE Pro by adding a proxy in front of it that has few simple "harness like features": - when the model gets stuck it tells it to "go on" - when it sends no output, malformed json, slips to wrong tool use format, etc it asks it to "try again better" - detects repetition and tells the model. - injects a prompt about "planning tool use" when it seems to be using same tools repeatedly. - injects a reminder it can use tools if there are no tool uses for over X messages.
10% - with just that.
I have more to test. My point is, open weights models get tested on naked model quality. "Frontier" models get tested as a model + whatever secret sauce they choose to put in front.
> At the heart of Multiverse's platform is CompactifAI, a compression technology that applies tensor networks, a mathematical framework from quantum physics, to the problem of AI model compression. This application was pioneered by co-founder and Chief Scientific Officer Dr. Román Orús and reduces the size of large language models by up to 80-95% with immaterial accuracy loss.
Ironic, considering they are releasing a 438B model that loses to a 27B one. From another part:
> Singularity Machine Learning is a cloud service that uses quantum machine learning for solving supervised learning problems.
I wouldn't be surprised if these guys just finetuned an open Chinese model and called it a day.
Easy enough to find out, ask it a whole bunch of questions about politically sensitive things that would be impossible to publish on CCTV, the Peoples Daily, CGTN, etc. If they didn't train the model and just fine tuned it, a lot of "don't talk about Tibet or the Dalai Lama or what happened in 1989" will be perma baked into it.
When they make being from Europe their whole personality, that shows how little value there is to the actual value proposition.
Do you even understand just how much Europe distrusts the US with its current administration at this point? Nobody wants to deal with US companies if they can.
I got the privilege to meet their CEO in Paris as they were announcing their partnership with Axelera and we exchanged some stories. They are full of very smart people, and they got accepted in the EIC Accelerator which, while you may not have heard of it, is one of the most rigorous and difficult programs to get into as a startup (far more difficult than YC which nowadays accepts any vague AI trash and always shoots wide).
EIC churns out DeepTech, invests 2.5M nondilutive with funding horizons of 10+ years and larger equity-based programmes too. Multiverse is a quantum computing company working on energy-efficient AI algorithms. The EIC is the reason companies such as ICEYE (https://www.iceye.com/) can exist. The chips running all that inference going in those US datacenters being built? 99% certainly european chips, almost certainly EIC-funded. (Eg. Hailo, Axelera, ..)
You haven't heard about the EIC because they don't care about marketing to americans; their target market is European companies, not you. But sorry to say, the talent is in Europe; the only reason you've been getting that talent in the US is a better business ecosystem, and your current administration has been utterly destroying this while our leaders have been actually fixing our issues. So keep shitting on Europe from the top of that crumbling hill, we'll see you when you emigrate.
Also, do you have a source for: "99% certainly european chips, almost certainly EIC-funded. (Eg. Hailo, Axelera, ..)"
For Hailo, the best I can find is https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-ai-a...
Which states that this is a 40 TOPS int4 PCIe device, not what you'd put in a datacenter and certainly not what can run a frontier class model (or the model above). Embedded devices for inference are really cool! But that's the opposite end of scale for what goes into a data center.
I remember doing my first bit of 'shader repurposing' a long time ago thinking 'wow, I wonder whether Intel realizes this graphics card is the more powerful processor in my computer'. Back then (in the pre-CUDA) days you had to contort yourself in the worst possible ways to get any kind of unintended computation done but when it worked it was amazing. Since those days NV has been pumping enormous effort into building out that miniscule head start that they had in the days of 'Brook' and the subsequent release of CUDA. It is going to cost a lot of money to duplicate that, and even more to compete with NV on what is now their home turf. But I really wished someone with very deep pockets would do just that.
A company, whose bread and butter is to remove parameters from models [1], releases a 438B model. Their wording is a little slippery, but they don't actually say they trained it.
I think this is GLM 5.2 with parameters removed. Its advertised in their changelog [2] as "capabilities are identical to GLM 5.2," it has the same two effort settings "high" and "max," and both are text only [3]. I might be wrong about this, but I would love to hear more about what they did before changing my mind.
Not the greatest fan of the marketing personally. Irrespective of what this model is.
[1]: https://multiversecomputing.com/compactifai/deployment
[2]: https://docs.compactif.ai/changelog/#added-3
[3]: https://docs.compactif.ai/features/multi-modality/#compatibi...
https://artificialanalysis.ai/models/quasar-438b
https://web.archive.org/web/20260901084009/https://artificia...
Despite my doubts I applied to one of their positions a couple of years back only to receive a super late and generic "we are not moving forward" mail (I'd say I fit pretty damn well for the position, but it seems to be the new normal that most companies don't even want to chat with people 99% of the time).
I'm sorry, what?
I don't get what the issue is. Chinese labs fully publish how they make great models. Architecture is known, training methods are often very open. Why Europe just can't copy what they do?
Their CMO's bio lists "20,000 'qualified' quantum AI contacts on linkedin" as the SECOND line in his bio.
"Business Management by ESADE. Co-Founder and CMO of Multiverse Computing. President, “barcelonaqbit-bqb”, 20,000 “qualified” quantum AI contacts on LinkedIn. VP of the AMETIC Innovation..."
Even if this fits the spirit of corruption, I'm not sure it fits the spirit of it. They certainly aren't there to build a product or provide valuable expertise, but taxes and salaries are paid, people go to conferences, they might even build something, but its a giant LARP. Don't expect to get rewarded or recognized for solid engineering.
I'm sure if you worked for a mid-sized Euro firm, you know what I'm talking about.
It seems Europe is many years behind in tech. Maybe for AI models they are just a decade behind, but as far as producing hardware capable to run SOA models they lag tens of years.