Open-Weights Models for Non-English Professional Writing

Every model card claims multilingual capability. The claims are true in the sense that the model will produce text in your language, and considerably less true in the sense that the text will have the register a professional context requires.

At small sizes, where local models live, language coverage is the most uneven capability there is. A 4B model that handles English business register comfortably may produce Dutch that is grammatically correct and reads like a translation. Here is how to find out before you commit.

Why Small Models Diverge On Language

Three reasons compound.

Training data proportion. English dominates the web, so English dominates training data. Other languages get a share that roughly tracks their online presence, which is not the same as their professional importance.

Capacity constraints. A 4B model has limited room. Multilingual capability competes with everything else for that room, and the compression falls hardest on languages with less data.

Register is subtler than grammar. Producing grammatically correct German is easier than producing German that a lawyer would send to a client. Formal register involves conventions, level of directness, and phrasing patterns that need substantial in-language professional text to learn. A model can be fluent and tonally wrong.

That third point is why “the model speaks my language” and “the model can write my work email” are different claims. Fluency is table stakes; register is the thing you actually need.

Rough Coverage Expectations

At the 1B to 8B tier, and this varies significantly by family:

Generally strong: English, Chinese, and usually Spanish, French, German, and Portuguese. Enough professional text exists in these that even small models pick up register.

Variable: Italian, Dutch, Polish, Russian, Japanese, Korean. Often fluent, sometimes tonally off. Worth testing rather than assuming.

Frequently weak at small scale: Nordic languages, Central and Eastern European languages, Arabic, Hindi, and most languages with a smaller online professional corpus. May be usable, may produce something that reads translated.

Larger models handle the second and third groups considerably better, which is the one place where the local-versus-cloud tradeoff genuinely bites for non-English writers.

Families Worth Trying First

Qwen has the broadest size range and strong multilingual training, with obvious emphasis on Chinese and English plus wide coverage elsewhere. Its 1.7B and 4B models under Apache 2.0 are the most practical starting point for local use, and it is the most downloaded open-weights family for a reason. Model cards are on Hugging Face.

Mistral has historically been strong on French, German, Spanish, and Italian, which is unsurprising for a European lab. If you write professional email in one of those, test it against Qwen rather than assuming the larger download wins. Release notes are at mistral.ai.

Gemma ships small Apache 2.0 models built for on-device use, with reasonable multilingual coverage and good prose sense at small scale. Documentation is at ai.google.dev/gemma.

None of these is reliably best for your language. Testing takes twenty minutes and beats reading three more comparison articles.

The Test That Actually Tells You

Public multilingual benchmarks measure translation and comprehension, not whether a model writes your professional register. Run this instead.

Take five real messages you have written in your language, including one formal message to a client or superior and one where you were being deliberately careful about tone.

Run each through your candidates with an instruction written in the target language, not in English. This matters more than people expect: an English instruction asking for German output frequently produces German that follows English conventions.

Then score four things:

Register: would you send this to the person it is addressed to, unchanged? Idiom: does it read as written by a native speaker, or as translated? Facts: did every name, date, figure, and commitment survive exactly? Length: same, shorter, or ballooned?

The second column is the one that fails most often and the one no benchmark measures. There is a fuller method in how to benchmark a local model on your own writing.

Write The Instruction In The Target Language

A concrete example of why this matters. For German business email:

Formuliere diesen Text als professionelle Geschäfts-E-Mail. Formeller Register, vollständige Sätze, Sie-Form. Beginne mit dem Anliegen und der Frist. Übernimm alle Namen, Daten, Zahlen und Zusagen unverändert. Füge keine Freundlichkeiten oder Inhalte hinzu, die nicht im Original stehen. Nicht länger als der Ausgangstext. Gib nur den umformulierten Text zurück.

That instruction specifies the formal address form, which is a German-specific requirement no English instruction expresses. Every language has equivalents: the tu and vous distinction in French, keigo levels in Japanese, formality registers that English simply does not encode.

Wrivio Contexts store the instruction per situation, so you write your language-specific version once and it applies to every rewrite. This is where Contexts earn the most for non-English writers: the conventions you need are hard to state and you only have to state them once.

The Fallback When Local Is Not Enough

If your language is in the third group and a local small model produces translated-sounding output, you have three options.

Use a larger local model. If you have a GPU, the 20B to 30B tier handles less-represented languages noticeably better. See how much VRAM you need for a local LLM in 2026.

Route by sensitivity. Local for anything confidential, accepting the register cost, and a frontier cloud model for non-confidential text where you want polish. The routing rule stays the same as always. See which tasks should stay local.

Use local for structure, edit the idiom yourself. A small model can reliably tighten structure and lead with the ask even when its phrasing is slightly off. You fix the idiom, which you can do faster than you could restructure. This is an underrated pattern and it keeps everything on your machine.

Common Questions

Do models write better in English than my language?

Almost certainly, at small sizes, unless your language is Chinese. The gap narrows with model size and it does not close.

Should I write my instruction in English or my target language?

Target language, for non-English output. It produces output following the target language’s conventions rather than English ones transposed.

Can a model handle mixed-language text?

Usually, and it is a reasonable place to state the expected output language explicitly. “Keep quoted English terms unchanged and write the rest in Dutch” is the kind of clause worth adding.

Is a translation model better than a rewriting model?

Different job. Translation moves content between languages; rewriting changes register within one. If you need both, do them as separate steps and check each, because combining them multiplies the ways facts can drift.

Download Wrivio for Windows to store a rewrite Context in your own language and run it locally.