agree. I have a usecase for a bigger open-weight model hosted by an EU company and the selection isnt really great, hard to make the regulatory gods (and the devs) happy at the same time right now
Good to see more developments in this space. I quite like this service, which is a little further than Hetzner and has several models to choose from: https://www.infomaniak.com/en/hosting/ai-services
Important note on Infomaniak's offering - when they first launched it, I gave it a try but despite claims that it's OpenAI API compatible, even basic things like sending a base64 encoded image were broken. I reached out to their support, who replied with (verbatim): "Unfortunately, we do not provide support for our AI Service, as the solution is highly unmanaged and uses our API."
"highly unmanaged" didn't fill me with confidence, and "uses our API" is a very weird reason to give for not offering support. I wrote off ever using them for anything beyond email after that reply.
I highly agree. I once registered a domain through them that happened to be my last name but coincidentally also the name of a company. They restricted me from using the domain and never refunded me the yearly fee for it (citing trademark reasons, although under Swiss trademark law, someones legal name is not a trademark violation, especially as it was just my personal website and I wasn't working on a competing product/service). Only weeks later someone else successfully registered it without such issues.
Worse even, in my account panel I still had control over the e-mail service of the domain, so there was most definitely a pretty bad security critical bug there.
Depends on your mindset. If you are an Engineer that does not need hand-holding and you accept terse answers from support with the gratitude to the human on the other side, then go for it.
You will at least not need months of training and tough exams to be an expert in Hetzner Cloud. It is simple, but that's the point.
I mostly use Hetzner baremetal servers and not cloud, but the support there is in my experience very competent, but also VERY german and direct. I can imagine if youre neither german or super technical that this can be intimidating or considered unfriendly
Interesting. I could see them perhaps coming in competitive for models that fit into single cards? Less so playing in the big model serving league...climbing into that esp right now would be madness
> The enable_thinking option is worth mentioning. Without it, the model can spend a surprising amount of the completion budget reasoning before it returns a visible answer.
Straight up the opposite, which the name makes abundantly clear, with the option it does reasoning, without it it doesn't...
llamacpp has a reasoning-budget and reasoning-message setting that can both be a global or per header setting. Using it, you can stop it's reasoning token count and insert a message at the point you stopped it.
This allows both the client and server to customize it. I typically use a message that tells it to compact the conversation and use subagents. I find the reasoning gets bloated when it's failed to do whatever task it's doing and often times it either has too little context (subagents) or its context is bloated (compact).
This works fairly well to get it to extend workable life up to ~1M on a local model.
This seems like a smart move, given their ability to host efficiently. I approve of efforts to make the cost of inference for smaller useful models slowly approach 'close to zero' and there are many good paths for getting there. It is useful for companies to get fast hosting for the class of smaller models they may end up hosting in house.
what is the hallucination here? it is fast, free and fun to try. And I genuinely think that the hardware decision (if they get bigger gpus) decides if this will be a banger product or not?
I read that data centers in Frankfurt use 40% of city energy, and seems that some fossil fuel based workarounds are being pushed. I hope we do better than the US in the energy and data center gold rush.
It would certainly be interesting to have a highly respected EU-native inference provider, if only to make the regulatory gods happy
Scaleway does that already but competition is always good.
agree. I have a usecase for a bigger open-weight model hosted by an EU company and the selection isnt really great, hard to make the regulatory gods (and the devs) happy at the same time right now
Scaleway have two separate (one fully managed one a bit less) services for that:
https://www.scaleway.com/en/generative-apis/
https://www.scaleway.com/en/inference/
yea, but no prompt caching right? This makes it unusable for my usecase at least, the cost would be insane
Good to see more developments in this space. I quite like this service, which is a little further than Hetzner and has several models to choose from: https://www.infomaniak.com/en/hosting/ai-services
Important note on Infomaniak's offering - when they first launched it, I gave it a try but despite claims that it's OpenAI API compatible, even basic things like sending a base64 encoded image were broken. I reached out to their support, who replied with (verbatim): "Unfortunately, we do not provide support for our AI Service, as the solution is highly unmanaged and uses our API."
"highly unmanaged" didn't fill me with confidence, and "uses our API" is a very weird reason to give for not offering support. I wrote off ever using them for anything beyond email after that reply.
Interesting, didn’t know about them! Weird model selection though, no glm or deepseek?
Infomaniak is such a shitty company, I had to use them on a previous job I worked at and dealing with them was awful.
I highly agree. I once registered a domain through them that happened to be my last name but coincidentally also the name of a company. They restricted me from using the domain and never refunded me the yearly fee for it (citing trademark reasons, although under Swiss trademark law, someones legal name is not a trademark violation, especially as it was just my personal website and I wasn't working on a competing product/service). Only weeks later someone else successfully registered it without such issues.
Worse even, in my account panel I still had control over the e-mail service of the domain, so there was most definitely a pretty bad security critical bug there.
That makes it sound like dealing with Hetzner is easy. Is the case?
Depends on your mindset. If you are an Engineer that does not need hand-holding and you accept terse answers from support with the gratitude to the human on the other side, then go for it.
You will at least not need months of training and tough exams to be an expert in Hetzner Cloud. It is simple, but that's the point.
I mostly use Hetzner baremetal servers and not cloud, but the support there is in my experience very competent, but also VERY german and direct. I can imagine if youre neither german or super technical that this can be intimidating or considered unfriendly
Being slightly on the spectrum, I find Hetzner's "VERY german and direct" support perfect.
I only have a couple of VPS plans with them, but my experience is good.
Interesting. I could see them perhaps coming in competitive for models that fit into single cards? Less so playing in the big model serving league...climbing into that esp right now would be madness
Why do you think this would be madness? It seems like, at least in the EU, there is barely competition for the big open-weight models?
> The enable_thinking option is worth mentioning. Without it, the model can spend a surprising amount of the completion budget reasoning before it returns a visible answer.
Straight up the opposite, which the name makes abundantly clear, with the option it does reasoning, without it it doesn't...
My writing wasnt clear here I think, without the option it defaults to reasoning enabled. With "without it" I meant without enable_thinking=False!
llamacpp has a reasoning-budget and reasoning-message setting that can both be a global or per header setting. Using it, you can stop it's reasoning token count and insert a message at the point you stopped it.
This allows both the client and server to customize it. I typically use a message that tells it to compact the conversation and use subagents. I find the reasoning gets bloated when it's failed to do whatever task it's doing and often times it either has too little context (subagents) or its context is bloated (compact).
This works fairly well to get it to extend workable life up to ~1M on a local model.
This seems like a smart move, given their ability to host efficiently. I approve of efforts to make the cost of inference for smaller useful models slowly approach 'close to zero' and there are many good paths for getting there. It is useful for companies to get fast hosting for the class of smaller models they may end up hosting in house.
I really hope they dont stop at the small models though! The bigger ones that dont fit on a single GPU are more interesting I think
Potentially interesting article ruined by AI slop hallucinations like
> For now, the API is fast, free, and fun to try. The next hardware announcement will tell us much more than another small model would.
what is the hallucination here? it is fast, free and fun to try. And I genuinely think that the hardware decision (if they get bigger gpus) decides if this will be a banger product or not?
nobody writes like that. the article is based on a prompt, which i would much rather read, instead of the ironed out version that the llm produced
fine, i agree that this doesnt sound like me. But the content is correct and not a hallucination!
Hetzner is very efficient hosting servers
Will this be the new division of labor?
Americans - best proprietary models
Chinese - best open weight models
Europeans - best / most efficient inference service
I read that data centers in Frankfurt use 40% of city energy, and seems that some fossil fuel based workarounds are being pushed. I hope we do better than the US in the energy and data center gold rush.
We need frontier open weight models from the West.
Maybe Nvidia's Nemotron can get there.
https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/
why?
Reasons include, but are not limited to:
a) broad competition is good
b) jurisdiction diversification (you don't want to be dependent on the regulatory winds of a single jurisdiction)
...but you're not dependent on anyone?
If I'm running a Kimi model on my local cluster, the CCP can't shut it off.
It's not like the Finnish government has veto power over Linux.