Build a “send as-is” button — a model that hands you the email, not advice about it

Look at what one team actually measured — and it stings. You ask an AI to write the email you've been putting off for three days. Back comes: "Of course! Here are three options", a preamble, three paragraphs, a note saying "pick the tone that fits". The finished text is in there — but it's buried. You have to dig it out, strip it down, and only then send it.

They counted how often models do this. By their measurement, Fable 5, GLM-5.3 and Kimi K3 bury the text in more than nine replies out of ten.

Then on 20 September 2026, open weights shipped for a model that just hands you the email. No options. No preamble. This isn't "writes nicer" — it's a different shape of answer: what comes back is the artifact itself, not a conversation about it. And that's the part you can build a button on.

Why this just became possible

The model is Hemmingway-1 from a team called Altworld: 27 billion parameters, Apache-2.0, built on Qwen3.8-27B, a 262,144-token context. Weights are open, commercial use included.

They trained it on exactly the writing you keep postponing: messages, emails, the awkward note to a colleague. By their own numbers it beats Fable 5.1 and leads GPT-6 Astra by fifty points on everyday requests, and in a blind "which of these two did a human write" test it finishes twenty-six points clear of the next model.

Straight talk about those numbers: three of the four benchmarks are their own, and they say so on the model card. Matchups were blind and judged by a different model — but don't take it on faith. The one independent result is the public EQ-Bench 4, run by its own harness: the model places third there, past GPT-5.5, Opus 4.7 and Opus 4.8.

For building, that's the whole point. As long as the reply comes wrapped in commentary, there's always a human with a mouse between the model and the send button. When the reply is the text, what stands between them is a button.

What you're building

A one-screen app:

1. One line about the situation. "Airline still hasn't refunded a cancelled flight after two months — I need a polite but firm email."
2. Three tone switches: warmer / drier / shorter.
3. The finished email and a copy button.

Why English: the model is english-first, and that lands right where it hurts. A support ticket to a foreign service, a refund chase, a note to a landlord abroad, a reply to a recruiter, an issue in someone else's repo — you want text that reads like a native wrote it, not like a translation and not like an AI. A translator translates text you already have; here there's no text yet. You describe the situation and get the email.

What you'll learn

• How to run an open model and call it from code. The repo gives you a one-line vllm serve, and quantized builds (GGUF and MLX) showed up the same day — which means it loads in LM Studio and serves a plain OpenAI-compatible API. No key, no bill.
• The difference between an assistant model and a tool model. You call the first to talk; you call the second to get a result. That decides whether your app still needs a human in the middle.
• What open weights buy you in practice. Nothing abstract here: the model shipped yesterday and it's running on your laptop today.

A starter prompt you can copy

The classic mistake is asking it to "write an email". Then the model invents facts and adds wrapping. A strong prompt pins the facts, the tone and the length — and explicitly bans everything except the text.

The weak version returns three drafts and a paragraph on choosing between them. The strong one returns a single text you can paste into your mail client and send.

The tone switches are then exactly one swappable line: Tone: warmer, apologetic / Tone: dry and short, no pleasantries / Length: under 60 words. Your whole "interface" is three buttons that change one line of the prompt.

What the result looks like

In: one line and three taps. Out: a hundred-word English email carrying your booking reference, the dates, a concrete ask, and not a single "I hope this email finds you well". You don't edit it — you send it.

And there's a side effect you notice on the third try: the scary email stops being scary. Not because something thought for you, but because the draft already exists and all you have to do is read it.

Honest limits, in the authors' own words:

• The model can be wrong and still sound certain — don't use it to decide anything medical, legal or financial. Drafting the complaint is fine; deciding what to do about the money isn't.
• It doesn't check facts. Reference numbers, dates and amounts come from you, in the prompt.
• A 27B model at four-bit quantization is roughly 16–17 GB. That's tight on a 16 GB laptop: take the MLX build on a Mac with 24 GB+, or start on the authors' own site (there's a free tier) and write the app against any OpenAI-compatible endpoint — you can move it local later.

A weekend plan

1. Friday evening. Describe, in one line, an email you're genuinely putting off. Run it and just look at the reply: is that the text, or a conversation about the text? That's the moment it clicks.
2. Saturday. Pull a quant into LM Studio (or run vllm serve), switch on the local server, and fire one request from code. Check that the reply contains nothing extra.
3. Sunday. One page: an input, three tone buttons, "copy". Add nothing else — the value is precisely that there are no steps.

Start with one email and one tone. Once you've sent the first one without editing a word, you'll know what to add next.

Source: Hemmingway-1 on Hugging Face