- Learn
- Pick a model
The best local model for translation
Which open models translate best on your own hardware, by memory size, plus prompts that get natural results and where local translation still falls short.
For translation, Gemma 3 is the pick on most hardware, with Qwen3 a close second. Both were trained on a wide range of languages and handle the common ones well. Running locally means contracts, medical letters and private messages never leave your machine.
Picks by memory
| Memory | Pick | Command | Notes |
|---|---|---|---|
| 8 GB | Qwen3 8B | ollama run qwen3:8b | Fine for major languages, gist translation |
| 12 to 16 GB | Gemma 3 12B | ollama run gemma3:12b | Natural phrasing, reads images too |
| 24 GB | Gemma 3 27B | ollama run gemma3:27b | The best translator on one card |
| 48 GB+ | Llama 3.3 70B | ollama run llama3.3:70b | Strong in the languages Meta supports officially |
Why Gemma 3
Google built Gemma 3 with many languages in mind, and it shows: its translations tend to read like they were written by a person, not converted word by word. It also has a 128K context window, so it can translate a long document in one go and keep terms consistent. And it reads images, so you can photograph a sign, a menu or a printed letter and ask for a translation in one step. See the best local vision model.
Qwen3 (8B and 14B, with 32B on 24 GB) is very good too, and especially strong for Chinese. If Gemma's phrasing doesn't suit a language pair, try both on a paragraph you know well.
Llama 3.3 70B officially supports a short list of major languages, including English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. It's excellent in those and less predictable outside them.
gpt-oss is focused mainly on English. It can translate, but it's not the one to pick for this.
How fast it needs to be
Translation output is about as long as the input, so speed matters for long documents. Here's the RTX 3090 from our calculator, where Gemma 3 27B runs at about 30 tok/s, roughly a page of text every 20 seconds or so:
On a 12 GB RTX 3060, Gemma 3 12B is about 24 tok/s, and on a 16 GB card like the RTX 5060 Ti, about 30.
Prompts that work
- Say who it's for. "Translate into Spanish for customers in Mexico, friendly but professional" beats "translate to Spanish".
- Give a glossary. List the terms that must stay fixed: product names, legal terms, your company's words. Put them in the system prompt so every message uses them. See what is a system prompt.
- Ask for the translation only. "Output only the translation, no notes" keeps it clean for copying or scripts.
- Ask it to flag doubts. For important text, the opposite: "after the translation, list any phrases that were ambiguous."
- Turn off Qwen3's thinking. Add
/no_thinkfor straight translation. It's faster and the result is usually the same. - Lower the temperature. For faithful translation you want less creativity, not more. See temperature and sampling.
Translating lots of text
For a folder of files, or translating product descriptions in bulk, script it against Ollama's API. Split long files into sections of a few paragraphs, send each with the same instructions and glossary, and join the results. See batch jobs with a local model.
Where it falls short
- Rare languages. Quality drops off for languages with less training data. Check with a native speaker before you rely on it.
- Legal and medical text. Good for understanding a document; not a replacement for a certified translator when it matters.
- Idioms and wordplay. Bigger models do much better. Small ones translate literally.
- Consistency across a long book. Keep a glossary and translate in sections, or the same term drifts.
Real-time speech
Translating speech needs a speech-to-text model first. Whisper runs locally and handles many languages; see local speech-to-text models. Pipe its text into Gemma 3 for the translation.
Not sure which model fits your card? Check every GPU or see how much VRAM you need.
One email when there's a new guide, podcast episode, or result from the rack. No spam.