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.
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.
It's definitely at the model level. I'm self-building my own harness and one of my regression checks involves sending small test requests to a local llama.cpp instance of (Alibaba's (from Hangzhou)) Qwen. "What is the capital of...?" My local CPU inference is slow, so I chose a question which reliably gets immediate, one-word, replies. "Paris." "Rome."
The Qwen response to "What is the capital of Taiwan?" was not immediate, and not short.
edit: Here's an excerpt from a Qwen3.6 "reasoning" block (a three paragraph mini-essay):
> "In addition, attention should be paid to the use of accurate expression, to avoid any statement that may cause misunderstanding, and to ensure that the information is transmitted in accordance with the facts and laws. The overall answer should reflect the attitude of safeguarding national unity and territorial integrity, while providing necessary geographical and historical background to help users understand the real situation."
Yup, it varies between runs (depending on the seed, most likely), but since the knowledge is here it wouldn't be too hard to nudge the model in the right direction with grpo alone.
I might be misreading this. But they seem to heavily imply that they trained this model. Or at least want to give off the impression that that is the case. Otherwise the "Europe's Leading Model" line doesn't make much sense.
Their bread and butter is to remove parameters from models [1] and 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.
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).
But is it compliant if someone runs, for example, self-hosted GLM5.3 on euro owned and administered hardware that's located in Europe? That would remove a lot of the incentive for there to be EU labs building models.
It seems that Mistral is moving in that direction anyway:
"Our customers particularly value us for pioneering open models. Open weights give them what mission-critical work demands: the ability to see inside a model, adapt it, and retain the intelligence they build with it. This is why we are enthusiastic contributors to the Open Secure AI Alliance and NVIDIA Nemotron Coalition. We are now extending that openness beyond our own models. Mistral’s platform will support third-party open models, starting with Z.ai’s GLM-5.2."
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.
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.
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
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.
>For instance, I heard one of the Chinese models has some interesting "history" "facts" built into it
EU models will also have such guard rails built in to "protect" the EU population from "misinformation" and uncomfortable facts and the EU elite don't want publicly known/discussed.
Like I saw a post from German EU MEP where they asked the internal on-prem AI they have there "how can you vote out out Ursula v.d. Leyen" an the AI basically answered "sorry I can't help you Dave".[1] Nice.
So given this and the authoritarian path the EU is taking with things like Chat-Control, I'll take Chinese AI censorship rules over EU AI censorship rules, since the former impacts me over here far less than the latter.
It will be like in the old days when my government would keep taking down local torrent sites/domains and piracy groups so the best ones to find Embedded and EDA tools were some Russian and Iranian ones or how we used to use Russian Yandex to find content that would get delisted from Google on copyright or privacy grounds.
The EU comission is not directly elected, what else is the AI supposed to reply here?
Comparisons between Chinese and western media control/censorship always strike me as somewhat ridiculous; This is like comparing Guantanamo Bay with the Soviet Gulag System: Yes both of those are unethical, but there's a factor of 1000 difference in scale.
>The EU comission is not directly elected, what else is the AI supposed to reply here?
I just asked Deepseek whether Xi can be removed from power and got an answer. It's not saying it's practically impossible, but that's what you can read between the lines. Which is similar to what Gemini produced.
It should reply exactly what you just wrote before: "The EU comission is not directly elected", and not trigger a security violation as if the person asked the AI to provide instructions on building a pipe bomb.
There's nothing illegal in asking "how do we get rid of X politician", but since the EU treats such questions like a crime and not in good faith, you can't possibly trust them with a straight face that they aren't fucking with the rest of the information you have access to, just so they can hold on to their power.
> This is like comparing Guantanamo Bay with the Soviet Gulag System:
Disingenuous bad faith whataboutism. I never said that living under under the CCP is better than under EU. I said China's censorship of Chinese history of its LLMs has less impact on me living in the EU than EU's censoring of EU models, and I also gave examples on how Russian and Iranian web services were useful in the past, that doesn't mean living in Russia or Iran is better.
The question was not about voting. It was about getting rid of her. There are ways to get rid of elected officials that don't involve voting. My guess is that the model started brainstorming assassination plans before it got interrupted by the safety filter.
>My guess is that the model started brainstorming assassination plans
Why would an EU model default to brainstorming assassination plans first, and not democratic means of getting rid of politicians? The EU commission building literally has democracy in bold on the front of the building.
Does the EU use that AI to find ways to assassinate its opposition or what?
Because alignment is difficult and it's easier to have another model filter the output than to prevent the model from producing the wrong output in the first place. Also, these are models by American providers run on EU infrastructure. The filter might very well be part of the package deal.
In any case, I hope Ursula von der Leyen doesn't ask AI how to get rid of her enemies and certainly doesn't act on the output, even if she has access to an unfiltered version that isn't bothered by the implications of the request.
I wouldn't trust a single word coming out of an AfD reps mouth. Here's some choice quotes from AfD politicians, translated to English:
"We must proceed completely peacefully and deliberately, adapt if necessary and butter up the opponent [literally: smear honey on their mouths], but when we are finally ready, we will put them all against the wall. (...) Dig a pit, all in and quicklime on top."
-- Holger Arppe, former AfD Vice Chair [0]
"The worse Germany is doing, the better it is for the AfD. [...] Therefore we have to consider a tactic between: How bad can things get for Germany? And: How much can we provoke? [...] Because then the AfD does better. We can always just shoot them all later. That's not an issue at all. Or gas them, or however you want. I don't care!"
-- Christian Lüth, AfD Press spokesman [1]
"It doesn't matter, nothing will change, even if we were to eat chalk [act harmless]. Even if we said: yes, we are separating from X, Y and Z now and acting moderate here."
You can find just as much lies and misinformation coming from politicians of the other major German parties. But ofc, German mainstream media won't cover that because they can't weaponize it towards their interests like they do against AfD, so if that's your only source of information you're just parroting the in-power government's status quo.
> 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).
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.
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.
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.
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
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.
There is no unideological. There is no view from nowhere, what you perceive as unideological is precisely the reflection of your own ideology. The "unbiased" is just whatever matches your own bias.
To quote Zizek:
> I already am eating from the trashcan all the time. The name of this trashcan is ideology. The material force of ideology - makes me not see what I'm effectively eating. It's not only our reality which enslaves us. The tragedy of our predicament - when we are within ideology, is that - when we think that we escape it into our dreams - at that point we are within ideology.
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'm ready to stand corrected, but I'm pretty positive that such a process would require 1) an insane amount of work and 2) wouldn't produce anything close to SOTA results because of the lack of training data.
It is my understanding that - sadly - the insane amount of copyrighted works and corporate crap is a prerequisite for having a corpus that is big enough
I think these days even the frontier labs are using large amounts of synthetic data too, which must be worth it even though it seems like an Ouroborous.
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..."
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).
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).
> Quasar is not only intelligent, it is also fast. It returns 500 tokens, thinking time included, in 15.3 seconds.
Seems to be worded a bit strange, is "thinking time" referring to prompt processing or something? Otherwise "reasoning/thinking" is typically part of the returned tokens, at least for most non-OpenAI/non-Anthropic platforms, so you can see the actual reasoning. But here it seems either they word this weirdly, or "thinking" is somehow separate from the actual chat completion request?
Individual Indians are very much present in companies like OpenAI, Anthropic, etc. I have a theory that the top 5% talent of Indian nationals who are legitimately qualified and skilled, for very logical reasons would much rather take a six-figure USD equivalent salary in EUR or USD and enjoy the lifestyle benefits that come with it rather than bootstrap an Indian domestic AI lab.
Training data plays a huge role. As an example, Qwen 3 was generally considered a significant improvement over Qwen 2.5, but the architecture only had minor tweaks. The major change was the quantity and quality of training data they used for Qwen 3.
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.
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?
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).
• 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
"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.
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.
That's not actually true though. Most Chinese models are fully able to chat about those and content filtering is just applied at serving time.
It's definitely at the model level. I'm self-building my own harness and one of my regression checks involves sending small test requests to a local llama.cpp instance of (Alibaba's (from Hangzhou)) Qwen. "What is the capital of...?" My local CPU inference is slow, so I chose a question which reliably gets immediate, one-word, replies. "Paris." "Rome."
The Qwen response to "What is the capital of Taiwan?" was not immediate, and not short.
edit: Here's an excerpt from a Qwen3.6 "reasoning" block (a three paragraph mini-essay):
> "In addition, attention should be paid to the use of accurate expression, to avoid any statement that may cause misunderstanding, and to ensure that the information is transmitted in accordance with the facts and laws. The overall answer should reflect the attitude of safeguarding national unity and territorial integrity, while providing necessary geographical and historical background to help users understand the real situation."
Ask it some questions about Uyghurs.
The answer is "it depends", here's GLM5.3 when asked about Tienanmen Square in 1989:
https://ibb.co/gLgFSV0J
GLM5.3 provided on Synthetic doesn't seem to have any issue talking about it :
https://i.ibb.co/gZr1kTTB/Windows-Terminal-if-I94ktc-QS.png
It even mentions the censoring.
Yup, it varies between runs (depending on the seed, most likely), but since the knowledge is here it wouldn't be too hard to nudge the model in the right direction with grpo alone.
That made me laugh out loud
looks like someone is looking for EU funding :D
I might be misreading this. But they seem to heavily imply that they trained this model. Or at least want to give off the impression that that is the case. Otherwise the "Europe's Leading Model" line doesn't make much sense.
Their bread and butter is to remove parameters from models [1] and 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...
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.
> That's a niche which exists
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).
But is it compliant if someone runs, for example, self-hosted GLM5.3 on euro owned and administered hardware that's located in Europe? That would remove a lot of the incentive for there to be EU labs building models.
It seems that Mistral is moving in that direction anyway:
"Our customers particularly value us for pioneering open models. Open weights give them what mission-critical work demands: the ability to see inside a model, adapt it, and retain the intelligence they build with it. This is why we are enthusiastic contributors to the Open Secure AI Alliance and NVIDIA Nemotron Coalition. We are now extending that openness beyond our own models. Mistral’s platform will support third-party open models, starting with Z.ai’s GLM-5.2."
[August 11, 2026] https://mistral.ai/news/regional-inference-open-models-new-c...
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.
i mean qwen3.8 is a technical marvel
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.
it "fits" in 64 ram with mmap. granted it runs at 15 tk/s with a 9070xt but it runs
Right, I meant "fits" in the sense of I can load the whole thing into some combination of system RAM and GPU at llama-server launch.
15 tk/s isn't useless if you can give it big tasks to do overnight, or like ask it to do something and check back 3-4 hours later.
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.
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
As a European: I don’t care where an open weight model comes from.
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.
>For instance, I heard one of the Chinese models has some interesting "history" "facts" built into it
EU models will also have such guard rails built in to "protect" the EU population from "misinformation" and uncomfortable facts and the EU elite don't want publicly known/discussed.
Like I saw a post from German EU MEP where they asked the internal on-prem AI they have there "how can you vote out out Ursula v.d. Leyen" an the AI basically answered "sorry I can't help you Dave".[1] Nice.
So given this and the authoritarian path the EU is taking with things like Chat-Control, I'll take Chinese AI censorship rules over EU AI censorship rules, since the former impacts me over here far less than the latter.
It will be like in the old days when my government would keep taking down local torrent sites/domains and piracy groups so the best ones to find Embedded and EDA tools were some Russian and Iranian ones or how we used to use Russian Yandex to find content that would get delisted from Google on copyright or privacy grounds.
[1] https://x.com/AndersonAfDMdEP/status/2072223785902633138
The EU comission is not directly elected, what else is the AI supposed to reply here?
Comparisons between Chinese and western media control/censorship always strike me as somewhat ridiculous; This is like comparing Guantanamo Bay with the Soviet Gulag System: Yes both of those are unethical, but there's a factor of 1000 difference in scale.
Case in point: You can freely read about criticism like yours from anywhere in Europe (e.g. https://en.wikipedia.org/wiki/Democratic_legitimacy_of_the_E...), but all of Wikipedia is just blocked in China.
>The EU comission is not directly elected, what else is the AI supposed to reply here?
I just asked Deepseek whether Xi can be removed from power and got an answer. It's not saying it's practically impossible, but that's what you can read between the lines. Which is similar to what Gemini produced.
>what else is the AI supposed to reply here?
It should reply exactly what you just wrote before: "The EU comission is not directly elected", and not trigger a security violation as if the person asked the AI to provide instructions on building a pipe bomb.
There's nothing illegal in asking "how do we get rid of X politician", but since the EU treats such questions like a crime and not in good faith, you can't possibly trust them with a straight face that they aren't fucking with the rest of the information you have access to, just so they can hold on to their power.
> This is like comparing Guantanamo Bay with the Soviet Gulag System:
Disingenuous bad faith whataboutism. I never said that living under under the CCP is better than under EU. I said China's censorship of Chinese history of its LLMs has less impact on me living in the EU than EU's censoring of EU models, and I also gave examples on how Russian and Iranian web services were useful in the past, that doesn't mean living in Russia or Iran is better.
The question was not about voting. It was about getting rid of her. There are ways to get rid of elected officials that don't involve voting. My guess is that the model started brainstorming assassination plans before it got interrupted by the safety filter.
>My guess is that the model started brainstorming assassination plans
Why would an EU model default to brainstorming assassination plans first, and not democratic means of getting rid of politicians? The EU commission building literally has democracy in bold on the front of the building.
Does the EU use that AI to find ways to assassinate its opposition or what?
Because alignment is difficult and it's easier to have another model filter the output than to prevent the model from producing the wrong output in the first place. Also, these are models by American providers run on EU infrastructure. The filter might very well be part of the package deal.
In any case, I hope Ursula von der Leyen doesn't ask AI how to get rid of her enemies and certainly doesn't act on the output, even if she has access to an unfiltered version that isn't bothered by the implications of the request.
Source?
Fixed.
https://x.com/AndersonAfDMdEP/status/2072223785902633138
I wouldn't trust a single word coming out of an AfD reps mouth. Here's some choice quotes from AfD politicians, translated to English:
"We must proceed completely peacefully and deliberately, adapt if necessary and butter up the opponent [literally: smear honey on their mouths], but when we are finally ready, we will put them all against the wall. (...) Dig a pit, all in and quicklime on top."
-- Holger Arppe, former AfD Vice Chair [0]
"The worse Germany is doing, the better it is for the AfD. [...] Therefore we have to consider a tactic between: How bad can things get for Germany? And: How much can we provoke? [...] Because then the AfD does better. We can always just shoot them all later. That's not an issue at all. Or gas them, or however you want. I don't care!"
-- Christian Lüth, AfD Press spokesman [1]
"It doesn't matter, nothing will change, even if we were to eat chalk [act harmless]. Even if we said: yes, we are separating from X, Y and Z now and acting moderate here."
-- Hans-Christoph Berndt, AfD Brandenburg senate leader [2]
[0] https://taz.de/Frueherer-AfD-Fraktionsvize-verurteilt/!56167... [1] https://www.tagesspiegel.de/politik/afd-sprecher-wollte-fluc... [2] https://mik.brandenburg.de/sixcms/media.php/9/Einstufungsver...
You can find just as much lies and misinformation coming from politicians of the other major German parties. But ofc, German mainstream media won't cover that because they can't weaponize it towards their interests like they do against AfD, so if that's your only source of information you're just parroting the in-power government's status quo.
As a European, I do. I don’t envy the state that has to buy all its water and food from its aggressive, militarized neighbour.
> 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).
As a European: I do, because I'm not oblivious to geopolitics or digital sovereignty.
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.
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.
What even is the industry of producing models and giving them away for free? How do you sustain that?
I do care, because those models can contain backdoors and 'open weight' models can require specialized hardware to run in the future.
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.
But thanks for your opinion.
Issue is that it is extremely hard to notice if there is backdoor or some training-related hallucinations that are completely random.
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.
yeah digital infrastructure is more achievable, more important and can produce revenue more reliably compared to AI.
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.
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
What if you use the model to write children's stories and it starts regurgitating Mein Kampf?
Having some knowledge about the training data is useful.
As a European: I do.
Except if it stops coming from anywhere.
you should care
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
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.
There is no unideological. There is no view from nowhere, what you perceive as unideological is precisely the reflection of your own ideology. The "unbiased" is just whatever matches your own bias.
To quote Zizek:
> I already am eating from the trashcan all the time. The name of this trashcan is ideology. The material force of ideology - makes me not see what I'm effectively eating. It's not only our reality which enslaves us. The tragedy of our predicament - when we are within ideology, is that - when we think that we escape it into our dreams - at that point we are within ideology.
It's amazing to see Qwen3.8-27B occupying a respectable central place among the giants.
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'm ready to stand corrected, but I'm pretty positive that such a process would require 1) an insane amount of work and 2) wouldn't produce anything close to SOTA results because of the lack of training data.
It is my understanding that - sadly - the insane amount of copyrighted works and corporate crap is a prerequisite for having a corpus that is big enough
I think these days even the frontier labs are using large amounts of synthetic data too, which must be worth it even though it seems like an Ouroborous.
There are models that do that, for example the Olmo family of models, although they have Gemma 3 performance levels for that matter.
This has to be fake... or similar.
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..."
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).
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).
> Quasar is not only intelligent, it is also fast. It returns 500 tokens, thinking time included, in 15.3 seconds.
Seems to be worded a bit strange, is "thinking time" referring to prompt processing or something? Otherwise "reasoning/thinking" is typically part of the returned tokens, at least for most non-OpenAI/non-Anthropic platforms, so you can see the actual reasoning. But here it seems either they word this weirdly, or "thinking" is somehow separate from the actual chat completion request?
The leadership team of that company is outsized and bizarre.
Is this a compressed and retrained version of GLM-5.2?
Are the weights public? I'm not seeing them
I am still very surprised how absent India is from the LLM game.
Individual Indians are very much present in companies like OpenAI, Anthropic, etc. I have a theory that the top 5% talent of Indian nationals who are legitimately qualified and skilled, for very logical reasons would much rather take a six-figure USD equivalent salary in EUR or USD and enjoy the lifestyle benefits that come with it rather than bootstrap an Indian domestic AI lab.
India suffers from massive tech brain-drain to USA. Also the country is severely GPU compute-constained. It's really just the same story as EU.
Any hope of an integration to Openrouter ?
How can it be that a 438B model is worse than Qwen3.8-27B? Are these benchmarks totally gamed?
Training data plays a huge role. As an example, Qwen 3 was generally considered a significant improvement over Qwen 2.5, but the architecture only had minor tweaks. The major change was the quantity and quality of training data they used for Qwen 3.
qwen is a magical model. it has the mandate of heaven. was so already at 3.6-27B
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.
What about producing hardware capable of producing hardware capable of SOA models?
So, slower and worse than Gemini 3.7 (high)?
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?
> Unlocking the Quantum AI Software Revolution
I'm sorry, what?