Hello,
I've been renting a small Kimsufi server for a few years to run a game for about ten players.
Last year I added a very small translation LLM (600 million parameters) that runs on the CPU.
The translation quality is obviously low, and I'm currently looking into whether it's possible to improve it.
OVH doesn't offer any cheap servers that also include a small GPU..
Hi,
I don't think so. It changes regularly and you can no longer see which servers are unavailable, so it's hard to answer. You can fall back on https://www.ovhcloud.com/fr/public-cloud/ai-endpoints/ or something else.
Yes, I saw this page. I admit I didn’t understand anything; it’s evolving so quickly.
Yes, okay, it’s an API service that OVH offers, similar to what Microsoft, Google, or Amazon provide in their clouds. But if I had wanted this solution I would already have used it last year.
I don’t want to depend on an external service because it requires maintenance, the API changes, the model is updated, and each time the client has to adapt their application. In short, it’s for large enterprises with an IT department.
Here, as an amateur, I want 100 % stability (no external dependency) and 0 % maintenance (because I don’t have time to take care of it).
So the solution of a small GPU on the machine would be ideal for me.
Ok, I completely understand you (and I think you’re right).
In the OVH lineup there’s a GPU filter:
https://www.ovhcloud.com/fr/bare-metal/prices/?display=list&gpu_brand=nvidia
Well, the first price is €1000 per month.
This filter doesn’t exist on https://www.kimsufi.com/fr/
We might need to look at CPUs that have an integrated GPU, maybe, but for running LLMs I have no idea what that will yield.
The subject is interesting—keep us posted 
No, €1,000 per month is out of the question, I pay that out of my own pocket.
I think for now I’ll leave what I have unchanged.
I’m using this model: https://huggingface.co/facebook/nllb-200-distilled-600M
with 600 million parameters, it enables translation between 200 languages and is open source (thanks Facebook). The quality is, let’s say, okay – except that in about 10 % of cases it misinterprets the sentence, but at least it’s helpful. It also has minor issues; sometimes it loops and outputs a phoneme that repeats forever, so you have to catch that in the script by killing the thread after 15 seconds. It translates a short sentence in roughly 3 seconds on CPU, which is fine.
So yesterday I tried a larger one, https://huggingface.co/facebook/nllb-200-3.3B
It’s the same base but with 3 billion parameters instead of 600 million. Translation now takes 20–30 seconds, which gets cumbersome. And the quality, um, I’ve had a few cases that were WORSE than the 600 M model, so I rolled back this morning.
I’m not good with LLMs or Python; the AI copilot wrote a Python script for me and I asked it for an interface that I could easily hook up to my game. After successive attempts and improvements we got it working. I did this last year during the holidays and it’s been running without maintenance for a year.
Copilot + last year... you should switch to a Claude Opus or a Codex, you might be surprised.
You have https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct
Apache 2.0 license, EU funding, but with 1.8 T parameters it might also be a bit slow. What’s your server?
Otherwise CTranslate2 is ultra‑fast, but you’ll need to assess the quality.
Do you have to create fresh translations all the time for players speaking different languages who are in the same game?
It's for translating the game's live chat. I mainly need French, English and German, but okay, players speaking other languages may show up.
CPU Intel Xeon E3-1245v2 - 4c/8t - 3.4 GHz/3.8 GHz, RAM 32 GB 1333 MHz
This is not my first attempt at adding a translation service; I've tried others before and it never succeeded, technical difficulties made me give up. And honestly I think few support 200 languages.
ok so even 5 seconds is still very slow for this kind of use case
https://github.com/OpenNMT/CTranslate2 will remain the fastest on a "small" machine (never tested, I don’t know the quality).
After that you’d still need to estimate external inference via an API endpoint; the cost might not be that high. You can run inference in the EU on OVH, Infomaniak, etc... OK you won’t be completely autonomous but at some point…
yes. On the other hand, maybe OVH will read this post and propose a GPU someday. I think that if a GPU (or an NPU) is available on the server side, we’ll find applications for it, not just translation. Microsoft pushed so hard to add an NPU to every PC but they forgot it’s unusable without loading a large model onto it. Otherwise, it’s also an option to explore, running this on client PCs.