ML Env

Use casesSelling across languages

Build your translation glossary from what you've already translated

Mines your existing translated content for the terms you already use, clusters the variants, and shows where you have been inconsistent.

How it works
It aligns your existing bilingual content — old catalogue exports, past agency deliveries, your current DE or FR site — page-to-page or segment-to-segment, extracts recurring product terms with the translations you actually used, and clusters variants with embeddings, so three different German renderings of 'drawstring' get grouped. The output is a candidate termbase that also shows where your wording has been inconsistent. A reviewed glossary makes the other translation tools on this page markedly more useful.
Data you need
A body of existing bilingual or multilingual content whose source and target texts can be paired: paired pages, old translation files, agency deliverables. A German site with no clear mapping to its English source yields much noisier results. If you have translated nothing yet, skip this and write a short glossary by hand instead.
What to expect
Extraction is noisy — expect to discard a fair share of candidate terms. It finds consistency, not correctness: where your past translations were themselves wrong, it will faithfully propose the wrong term. If your old translations were free rewrites rather than close translations, alignment quality drops and so does the candidate list; working only from the live site because the agency files are long gone still works, but needs more manual clean-up.
Where people stay involved
A person — ideally one fluent speaker per market — picks the winning translation for each clustered term and marks do-not-translate entries. The model surfaces candidates and inconsistencies; it does not decide your terminology.

Which model, and what it costs to run

Qwen3-8B

No public benchmark ranks this particular job, so this is not a leaderboard pick — it is where we would start: small, permissively licensed, and strong wherever it has been measured. We prove it on your own content before anything ships.

Licence
Apache-2.0
Weights at 4-bit
5 GB
Context
32K tokens
Publisher
Alibaba (Qwen Team)

What the hardware costs

One 48 GB card holds it

Rent in the EU
$1.60/hrScaleway, Paris (PAR2)
Buy the card
$7,569new, one-off
Or rent it by the token
$0.04 / $0.04per M in / out · DeepInfra

Hardware only, third-party prices from 2026-07. The figure excludes the KV cache, which grows with context length and how many people use it at once — sized properly in a conversation, not guessed here. Renting by the token is cheaper up front; why our customers still self-host is below.

European languages

openGPT-X European LLM Leaderboard · FLORES-200 translation, BLEU · 10 models

Previous-generation models — the board has not been rerun on the 2026 roster, so read the order of magnitude, not a ranking of today's models. And BLEU scores adequacy, not tone: it cannot tell you whether the register fits your brand. That is judged on your own copy.

EuroLLM-9B25.6 · EULlama 3.1 70B24.3Mixtral 8x7B17.3Mistral-Nemo 12B17.1Teuken-7B sigma17 · EULlama 3.1 8B17Llama 3 8B15.9Teuken-7B v0.615 · EUCommand-R 35B13.8Mistral-7B v0.313.3
Higher BLEU is closer to a human translationEU-sovereign model

Independent measurement · openGPT-X / Fraunhofer IAIS · captured 2026-07-10

The API is cheaper per token. Here is why our customers don't use it.

We will not pretend otherwise: renting a model by the token from a serverless API costs less per million tokens than a card we run for you. We show that price on every use-case page. What it does not include is the part a shop with a customer database actually pays for.

  1. 01

    Your data never leaves hardware you can point at

    A serverless "we don't retain your data" is a clause in a contract. Running the model on a card in Amsterdam is a fact of architecture: your catalogue, tickets and customer records are never sent to a third party at all. For a GDPR audit, that is the difference between a promise and a floor plan.

  2. 02

    The price cannot move without your say-so

    A serverless rate card is somebody else's lever. The provider can raise the price, retire the model, or change the terms, and your cost moves with it. The same model on the same card costs the same next year — you own the number.

  3. 03

    The model cannot be taken away

    Hosted APIs deprecate models on their own schedule; the one you built on can be gone in a quarter. An open-weight model on your own hardware runs for as long as you keep the lights on. No vendor can end-of-life it out from under you.

And the price gap closes with volume: past a card you keep busy — very roughly four billion tokens a month — owning is cheaper outright, even before the three reasons above.

Getting a case like this one from a conversation to production takes about two months, and you can stop at the end of any phase.

Next

Tell us what your team does by hand

Describe the process that takes the most time. We will say plainly whether a model is the right tool for it.