作者： Jakub Rusinowski · 最后更新： 2026年7月10日
A local LLM for writing emails drafts, replies, and summarizes on your own computer — no cloud API, no subscription, and nothing you type ever leaves your machine. Any modern 7–8B model running in Oll
A local LLM for writing emails drafts, replies, and summarizes on your own computer — no cloud API, no subscription, and nothing you type ever leaves your machine. Any modern 7–8B model running in Ollama or LM Studio handles email and everyday writing well, and this guide shows the exact setup and the prompt patterns that make it genuinely useful day to day.
Last Updated: July 2026
For most tasks, cloud models are simply smarter. Email and short-form writing is the exception where local wins outright, for three concrete reasons:
The honest tradeoff: a 7–8B model won't research facts for you and can mangle very long threads. For drafting, replying, tone-shifting, and summarizing — the 95% case — it's more than enough.
Install Ollama (or LM Studio if you prefer a GUI — see our Ollama vs LM Studio comparison), then pull a model:
# Best all-round writer at 8B — needs ~6 GB VRAM or a 16GB Mac
ollama run llama3.1
# Alternative with a slightly more natural, less "assistant-y" tone
ollama run gemma3:12b
# For 8 GB machines / older laptops
ollama run qwen2.5:7bAny of these runs on an 8–12 GB GPU or a 16 GB Apple Silicon Mac. Not sure what your machine handles? Check it in the hardware analyzer.
The difference between a useless local model and a great one is almost entirely the prompt. These three templates cover nearly all email work.
Give the model the facts and the tone; never make it invent content:
> Draft an email from these points. Tone: friendly but professional. Keep it under 120 words. > - can't make Thursday's call > - propose Monday or Tuesday afternoon instead > - the Q3 numbers they asked about are attached
Small models excel here because you supplied every fact — the model only supplies the connective tissue.
> Below is an email I received. Write a reply that: declines the offer politely, leaves the door open for next quarter, and thanks them for the detailed proposal. Keep my reply shorter than their email. > > [paste the email]
The instruction list *before* the pasted email matters — 7–8B models follow instructions better when they come first.
> Summarize this email thread in 3 bullet points, then list any action items with who owns each one. Flag any deadline mentioned. > > [paste the thread]
For very long threads, mind your context window — an 8B model at default Ollama settings reads about 2,000 words reliably. Longer than that, raise the context size (see our context windows guide) or summarize in chunks.
Instead of repeating "professional but warm" in every prompt, bake your voice into the model with a system prompt. In Ollama, create a Modelfile:
FROM llama3.1
SYSTEM You are my email writing assistant. Write in a direct, warm, professional voice. No corporate filler phrases ("I hope this email finds you well", "per my last email"). Short paragraphs. Never invent facts I did not provide — if information is missing, put [PLACEHOLDER] instead.ollama create email-assistant -f Modelfile
ollama run email-assistantThe [PLACEHOLDER] instruction is the single highest-value line: it stops the model from hallucinating meeting dates and figures you never mentioned. For deeper system-prompt technique and reusable personas, see System Prompts 101.
| Model | VRAM (Q4) | Writing style | Best for |
|---|---|---|---|
| Llama 3.1 8B | ~6 GB | Balanced, reliable instruction-following | Default choice |
| Gemma 3 12B | ~8 GB | Most natural prose, least robotic | Client-facing email |
| Qwen 2.5 7B | ~5 GB | Concise, slightly formal | Low-VRAM machines |
| Mistral NeMo 12B | ~8 GB | Good long-thread summaries (128K context) | Digest-style work |
All four handle drafting and replying well; the differences show up in tone. If your emails read stiff, switch models before you blame your prompts.
1. Keep a terminal (or LM Studio window) open with your email-assistant model loaded — an 8B model stays resident in ~6 GB and responds instantly. 2. Copy the email you're answering, paste it under a Pattern 2 instruction block. 3. Read the draft critically. You are the editor; the model is the intern. Fix the one awkward sentence, then send. 4. For anything sensitive — HR, legal, medical — this is the *only* AI workflow you should be using at all. That data has no business in a cloud API. For business/GDPR context, see Private AI for Business.
Yes. Email drafting is a low-difficulty task: you provide the facts, the model provides fluent phrasing. A 7–8B model like Llama 3.1 handles it indistinguishably from cloud models for the large majority of messages. The gap only appears in long, complex, multi-document work.
Llama 3.1 8B is the safest default; Gemma 3 12B produces the most natural-sounding prose if you have ~8 GB of VRAM or a 16GB+ Mac. Avoid going below 7B for writing — 3B-class models produce noticeably stilted text.
No. An 8B quantized model runs at usable speed (10–20 tokens/sec) on a modern CPU with 16 GB RAM, and very well on any Apple Silicon Mac. A GPU makes it instant rather than merely fine.
Not out of the box — Ollama and LM Studio are chat tools, not email clients. The copy-paste workflow above is the simple, safe default. If you want automation, both expose a local API you can script against — see Setting up a Local API.