Mistral Small 3.2, Qwen3.8-27B and Gemma 4 12B cover most business writing
Sizes are Ollama download sizes at the default quantization; context memory is extra. Figures marked developer-reported come from the model card or the Ollama library page, not from a PromptQuorum test.
- Mistral Small 3.2 24B (`ollama run mistral-small3.2`, 15 GB, 128K context, Apache 2.0, Mistral AI, France): best for emails and memos. Mistral reports better instruction following and about half the repetition loops of Small 3.1 (1.29% vs 2.11% infinite generations, developer-reported). It replaces Mistral Small 3.1, which earlier versions of this page recommended.
- Qwen3.8-27B (`ollama run qwen3.8:27b`, 18 GB download, 256K context, Apache 2.0, vision-language): best for long proposals, reports and contract summaries, and it can read screenshots or scanned pages. Plan for about 24 GB of memory in use (see our best local LLMs ranking).
- Gemma 4 12B (`ollama run gemma4:12b`, 7.7-8.0 GB, 256K context, Apache 2.0): best for 12-16 GB machines and quick internal messages. Google reports 83.4% MMMLU for the 12B model.
- Gemma 4 31B (`ollama run gemma4:31b`, 19-20 GB, 256K context): the strongest multilingual model in the Gemma 4 family at 88.4% MMMLU (Google-reported); fits a 24 GB GPU.
- Qwen3.5 9B (`ollama run qwen3.5:9b`, 6.6-7.6 GB, 256K context): best for 8 GB machines and non-English business email; the Ollama library lists 201 languages and dialects.
- Ministral 3 8B (`ollama run ministral-3:8b`, 6.0 GB, 256K context, Apache 2.0, Mistral AI): an 8 GB alternative with dozens of languages including English, French, Spanish, German, Italian, Portuguese, Chinese, Japanese, Korean and Arabic; it needs a recent Ollama (0.13.1 or newer).
- Superseded picks: Mistral Small 3.1, Qwen3 8B and Llama 3.1 8B are still downloadable but predate the 2026 generation above; a Llama 3.3 70B is no longer needed for long documents because Qwen3.8-27B offers 256K context at a fraction of the memory.
Match the model to the writing task
Task | Best model | Prompt tip |
|---|---|---|
| Task | Best model | Prompt tip |
| Email drafting | Mistral Small 3.2 24B | "Active voice, max 150 words, no jargon" |
| Executive memo | Mistral Small 3.2 24B | "Format: Problem / Recommendation / Next steps" |
| Business proposal (1-3 pages) | Qwen3.8-27B | Paste 2-3 past proposals as style reference |
| Contract or report summary | Qwen3.8-27B | "List obligations and risks, quote the clause" |
| Slack or internal message | Gemma 4 12B | "Casual but professional, 2-3 sentences" |
| Non-English business email | Qwen3.5 9B | "[Language] business email, formal register" |
| Polishing an existing draft | Any model on this page | "Edit for clarity, cut buzzwords, active voice" |
Brand voice comes from examples, not from model size
Business writing needs consistency. Teach the model your voice with real text.
- 1Gather examples: 3-5 emails or memos in your brand voice. Use real sent emails, not idealized ones.
- 2Analyze the style: ask the model to describe sentence length, formality, vocabulary, openings and closings, then turn that into a one-paragraph style guide.
- 3Create a prompt template: "You write like this: [EXAMPLES]. Now draft [TASK] in this voice."
- 4Specify constraints: "Keep to 150 words.", "Use active voice.", "No jargon or buzzwords."
- 5Store templates: save one prompt per writing type (sales, support, internal) and reuse it.
Five mistakes make local drafts worse
- Choosing a bigger model than the task needs. For emails and memos a 12B-27B model is enough; a larger one mostly adds memory use and waiting time.
- No examples provided. The model guesses your voice. Always give 3-5 real sent emails or memos.
- Trusting the first draft. Business writing needs one or two edit cycles; use edit prompts, not generation only.
- Ignoring context length: Ollama picks its default context from available memory (4K tokens under 24 GiB, 32K at 24-48 GiB, 256K above), so a long proposal can be cut off on a smaller machine. Raise it in the Ollama app settings or with `OLLAMA_CONTEXT_LENGTH=32768` when starting the server, and check `ollama ps`.
- Using one model for everything: draft with a fast model such as Gemma 4 12B, then polish with Mistral Small 3.2 24B in edit mode to tighten tone and remove jargon.
Set up a local writing assistant in five steps
- 1Install Ollama (see the install guide) and open a terminal.
- 2Pull the model that fits your memory, for example `ollama run mistral-small3.2` for a 24 GB machine or `ollama run gemma4:12b` for a 16 GB laptop.
- 3Write a system prompt with 3-5 examples of your brand voice and save it in your chat app or an Ollama Modelfile.
- 4Raise the context length for long documents (`OLLAMA_CONTEXT_LENGTH=32768`) and check it with `ollama ps`.
- 5Paste your draft and ask: "Refine this email for [tone], max 150 words." Review the result before sending.
Regional context: EU, Japan and Germany
EU / GDPR
When business emails contain names, contact details or deal terms, running the model locally means that text is not sent to a cloud AI provider, which typically removes the need for a data processing agreement (GDPR Article 28) with that provider for the writing step. Your own GDPR duties for storing and using the drafts remain. This is general information, not legal advice. Mistral Small 3.2 24B (Mistral AI, France, Apache 2.0) is an EU-origin option with a permissive licence.
Japan
Japanese business writing uses strict politeness levels (teineigo, sonkeigo, kenjōgo). Models default to polite speech but often mix registers unless instructed. Add an explicit instruction such as 「メールは丁寧な敬語(尊敬語と謙譲語)を使用してください」 and review the first draft; Qwen3.5 9B and Gemma 4 are reasonable starting points.
Germany
German business correspondence uses Sie-form, full company names and a factual tone. Add to your system prompt: "Schreiben Sie auf Deutsch mit Sie-Anrede, sachlichem Ton, ohne Anglizismen." Mistral Small 3.2 24B and Gemma 4 follow such instructions well in our editorial judgment; check the first draft for du-form slips.
Frequently Asked Questions
What is the best local LLM for writing emails?
Mistral Small 3.2 24B (`ollama run mistral-small3.2`, 15 GB download) is our pick for emails and memos: Mistral reports stronger instruction following and fewer repetition loops than Mistral Small 3.1. On a 16 GB laptop use Gemma 4 12B (`ollama run gemma4:12b`); on 8 GB use Qwen3.5 9B (`ollama run qwen3.5:9b`).
What is the best local LLM for writing in general?
For business writing (emails, memos, proposals, reports) use the picks on this page: Mistral Small 3.2 24B, Qwen3.8-27B or Gemma 4 12B depending on your memory. For fiction and poetry, see the separate creative writing ranking.
Which local LLM is best for a small team?
Run one model on a shared workstation with 24 GB or more of memory: Qwen3.8-27B (18 GB) for long documents or Mistral Small 3.2 24B (15 GB) for short business text. Team members then reach it through a chat front end instead of each installing a model.
Do I need a 70B model for business writing?
No. In our editorial judgment, a 12B-27B model covers emails, memos and proposals; a 70B model mainly adds memory use and waiting time for short business text. Larger models help most with very long, multi-section documents.
How do I make a local LLM match my brand voice?
Paste 3-5 real sent emails into the system prompt and say "Write in this style". To analyze your style first, ask: "Describe the sentence length, formality, vocabulary, openings and closings of these emails, then write a one-paragraph style guide." Reuse that guide as the system prompt and refresh the examples every six months.
Can local LLMs write in German Sie-form?
Yes, with an explicit instruction: "Schreiben Sie auf Deutsch. Verwenden Sie ausschließlich die Sie-Form. Sachlicher, professioneller Ton ohne Anglizismen." Without it, a model can drift into du-form, so always put the instruction in the system prompt and check the first draft.
What is the best local LLM for non-English business writing?
On 8 GB machines, Qwen3.5 9B (201 languages and dialects per the Ollama library) is our pick. On a 24 GB GPU, Gemma 4 31B has the highest multilingual score in Google's Gemma 4 model card (88.4% MMMLU, developer-reported). For formal European languages, Mistral Small 3.2 24B is also a strong choice.
Is it better to edit a draft or generate from scratch?
Editing is safer and needs less prompting: give the model your draft and ask it to cut buzzwords, shorten sentences and keep your tone. For complex documents, draft with Gemma 4 12B, then polish with Mistral Small 3.2 24B in edit mode.
Can I use a local LLM for confidential emails?
Yes. The model runs on your machine, so the text is not sent to an AI provider. Your own data-protection duties (for example GDPR) still apply to how you store and use the drafts. This is general information, not legal advice.
How do I avoid output that sounds like generic AI text?
Provide real brand examples, ask for active voice and short sentences, and ban filler phrases. If the result is too formal, add: "Remove buzzwords. Use everyday language. Write like you are messaging a colleague."
Sources
- Mistral AI. "Mistral-Small-3.2-24B-Instruct-2506" model card. huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506 -- Apache 2.0 licence, Small 3.1 vs 3.2 benchmarks (developer-reported).
- Qwen Team. "Qwen3.8-27B" model card. huggingface.co/Qwen/Qwen3.8-27B -- Apache 2.0 licence, 262,144-token native context, vision-language model.
- Google. "Gemma 4 model card". ai.google.dev/gemma/docs/core/model_card_4 -- sizes, context windows, Apache 2.0 licence, MMMLU scores (developer-reported).
- Ollama library pages: mistral-small3.2, qwen3.8, gemma4, qwen3.5, ministral-3 -- tags, download sizes and context windows, checked 2026-11-19.
- Ollama documentation. "Context length". docs.ollama.com/context-length -- default context by available memory.
