Use case 12 of 13

Private translation for documents and speech

Translate contracts, emails and recordings between dozens of languages on your own hardware.

Paper sheets with a blue paperclip, blue headphones and a black mini PC — Image: generated with Grok Imagine

The real example

LibreTranslate, and a new generation of translation LLMs

LibreTranslate is a free, open-source machine-translation API (AGPL-3.0) that is entirely self-hosted and offline-capable, powered by the Argos Translate library. Newer translation-specialised LLMs raise the bar: Tencent's Hunyuan-MT 7B supports mutual translation among 33 languages, and its card says it placed first in 30 of the 31 language categories it entered at WMT25. For speech, Whisper large-v3 transcribes many languages, and Parakeet TDT v3 covers 25 European languages.

Sources: LibreTranslate on GitHub · Hunyuan-MT 7B model card · Parakeet TDT 0.6B v3 model card

Why run it locally

Legal, medical and HR documents are often exactly what you must not paste into a web translator.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Silent home server

$1,283Sample

16 GB on GeForce RTX 4060 Ti 16 GB

Hunyuan-MT 7B · 4.9 GB (~4-bit)
Whisper large-v3 turbo · 1.6 GB (16-bit)

Weights ≈ 6.5 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Argos models are tiny; the 7B LLM handles nuance.

Runtime: LibreTranslate for quick jobs; llama.cpp for the LLM

Mid

Used-3090 24 GB

$1,980Sample

24 GB on GeForce RTX 3090 24 GB (used)

Hunyuan-MT 7B · 4.9 GB (~4-bit)
Qwen3 14B · 9.0 GB (Q4_K_M GGUF file)
Whisper large-v3 · 3.1 GB (16-bit)

Weights ≈ 17.0 GB of 23.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Use a 14B general model as a second opinion on hard passages, with Whisper loaded for recordings.

Runtime: llama.cpp server + a small batch script

Pro

Strix Halo SFF 128 GB

$3,499.99Sourced

128 GB on Ryzen AI Max+ 395 mini PC 128 GB

gpt-oss-120b · 63 GB (MXFP4 GGUF file)
Hunyuan-MT 7B · 4.9 GB (~4-bit)
Whisper large-v3 · 3.1 GB (16-bit)

Weights ≈ 71.3 GB of 96.0 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Translate long documents in one pass with a large context.

Runtime: llama.cpp

Setup outline

Run LibreTranslate with Docker for a quick, offline API.
For higher quality, serve a translation LLM with llama.cpp and send paragraphs with a fixed prompt.
Keep a glossary of names and terms and include it in every prompt.
For recordings, transcribe with Whisper first, then translate the text.

Limits

Hunyuan-MT's licence is not declared in Hugging Face metadata. Read the repository licence before commercial use.
Low-resource languages remain weak across the board.
Machine translation of legal text still needs a qualified human check.

Next: Personal and small-business finance analysis

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.