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.
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.
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.
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.
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.
> 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?
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 someone working at the EC 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". Nice.
So given this and the authoritarian path the EU is taking with things like Chat-Control, I'll take Chinese censorship rules over EU censorship rules, since the former impacts me over here far elss the latter.
> 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).
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
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.
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.
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.
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).
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.
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.
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.
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.
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.
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.
> 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.
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.
> 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?
As a European: I don’t care where an open weight model comes from.
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 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 someone working at the EC 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". Nice.
So given this and the authoritarian path the EU is taking with things like Chat-Control, I'll take Chinese censorship rules over EU censorship rules, since the former impacts me over here far elss the latter.
Source?
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.
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
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.
But thanks for your opinion.
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.
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.
Except if it stops coming from anywhere.
As a European: I do.
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.
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).
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.
Is this a compressed and retrained version of GLM-5.2?
Are the weights public? I'm not seeing them
The leadership team of that company is outsized and bizarre.
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
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 ?