Choosing a model

Browser engine — the six models

These run inside Chrome on your GPU. The numbers come from the model list the extension actually ships, not from estimates. Two VRAM columns, because which one applies to you depends on your GPU — see the note underneath.

Which VRAM column applies to you? The models are built for f16 shaders. Some GPU drivers expose WebGPU without the shader-f16 feature — Chrome on Linux today is the common case — and there the f16 build fails to compile. emailito detects this and automatically substitutes the f32 build of the same model: identical quality, roughly 15–25% more VRAM and a larger download. The panel tells you when it makes the switch, so you are never guessing.

Does the task change which model you want?

Yes, more than the size difference suggests.

Short replies are forgiving. A 3B model handles “thanks, Tuesday works, I will send the file tomorrow” perfectly well.

Summaries are not. A summary has to stay faithful to facts spread across a long thread — who owes what, which date was agreed, what the deposit was. Small models blur exactly that kind of detail, and a summary that quietly invents a date is worse than no summary. If you mostly use Summarize thread, the 7B is the one to run.

Long threads favour bigger models for the same reason. Every engine caps the thread it sends at roughly 24,000 characters — emailito keeps the first message and as many of the most recent as fit, and marks the gap — so on a very long thread the model is reasoning from an abridged version either way. A larger model does more with that abridgement.

Mail that is not in English is where the gap is widest. Qwen 2.5 covers far more languages than Llama 3.2's official eight, which is why the default is a Qwen model at both 3B and 7B.

How much VRAM do I actually have?

• Any OSOpen chrome://gpu and look for “GPU0” — it names the adapter, and whether WebGPU is enabled at all.
• WindowsTask Manager → Performance → GPU. “Dedicated GPU memory” is the number that matters.
• macOSApple Silicon shares one pool of memory between CPU and GPU, so there is no separate VRAM figure. A 16 GB M-series machine runs the 7B comfortably.
• LinuxRun nvidia-smi, or glxinfo | grep -i memory on other drivers.

Free VRAM matters, not total. Other tabs, an external display and the desktop compositor all take a share. If a model fails to load on a card that looks big enough on paper, close some tabs and try again before dropping a size.

Languages

emailito writes in the language of the mail you are answering — a Spanish thread gets a Spanish draft, and summaries follow the thread's language too. There is no setting for it: the model is told to match, every time.

How well it writes that language is a model-size question, and a sharper one than for English. Qwen 2.5 is the default browser model partly for this reason — Llama 3.2 officially covers eight languages, Qwen 2.5 covers far more. English is the most reliable, and Spanish and French read well in our own testing. Italian and German are understandable on the 3B default but drift: a word from the wrong language slips into an otherwise fluent draft, and details can come back wrong. On a 7B/8B, Ollama or Claude both are clean. Languages we have not tested may be rougher still, so try your own before relying on it. Summaries hold up better than replies everywhere, because a summary recombines wording already in the thread while a reply has to invent sentences from scratch.

The first run

The first draft on a new model downloads its weights — about 2 GB for the 3B default, around 5 GB for the 7B — and the browser caches them. After that the model loads from cache in seconds and works with no network at all. You do not have to wait for it mid-draft: settings has a Download model now button that fetches and loads the weights while you are already there expecting a wait, a Test browser engine button that reports whether your GPU can run it and whether it is already downloaded, and a Remove download button that deletes the weights again when you want the disk space back.

Weights are downloaded from huggingface.co. They are data files, not code — the executable part of the engine ships inside the extension package, which is why emailito can honestly answer “no” to the Chrome Web Store's remote-code question.

Ollama, if you already run it

Ollama uses your GPU natively rather than through WebGPU, so it fits larger models than the Browser engine and works on machines where WebGPU does not. emailito talks to it over loopback only and refuses any non-localhost URL, so this stays on your machine.

Avoid reasoning models such as qwen3 here. They either think for a long time before answering or leak their reasoning into the draft. A plain instruct model answers directly, which is what drafting needs.

Claude, when quality matters more than locality

Claude writes better email than anything you can run locally, and it is the one mode where your email leaves your machine. It goes straight from your browser to Anthropic with your own API key — there is no emailito server in the path — and you approve an explicit disclosure before the first request. Sonnet 5 is the default; Haiku is cheaper and quicker, Opus is the strongest. All three need API credit at console.anthropic.com, which is a separate balance from a Claude.ai subscription.

Still not sure?

Run the default for a week. If drafts need more editing than they save, move up a size before giving up on on-device entirely — most disappointment with local models is a model-size problem, not a local-versus-cloud one. Questions and odd behaviour are welcome in the issue tracker.