Gemini is Google's general AI assistant and its family of models. It answers questions, writes, summarises, helps with code and works inside many Google products. Hemmingway is narrower on purpose: it is built for writing to people, and for making that writing sound like you wrote it.
One thing first: we did not benchmark Gemini. None of our tests included a Gemini model, so this page has no scores for it and makes no claim about which writes better. It compares the two on what each is built for, how you use them, privacy and open weights, and it shows you how to test them side by side yourself.
The short version
| Hemmingway | Google Gemini | |
|---|---|---|
| Built for | Writing that sounds like a person | A wide range of tasks |
| Model | Hemmingway-1, 27B, our own | Google's Gemini models |
| Open weights | Yes, Apache-2.0 | Not the Gemini models themselves |
| Where you use it | Apps for Mac, Windows and Android, the web, an API | Google's apps, the web and Google products |
| Drafts replies from your inbox | Yes, in the apps | Not covered here |
| Trains on your chats | Never | See Google's own settings |
| Tested by us | Yes | No |
What each is built for
Gemini is a general assistant. You can ask it about a recipe, a spreadsheet formula, a news story or an email, and it will have a go at all of them. For many people that breadth is the point.
Hemmingway does one job. Hemmingway-1 was trained on the writing people actually do every day: messages, emails, the awkward note to a colleague, the thing you have been putting off. It was trained to hand you the message itself, one version, with no preamble and no list of options, in the register the situation calls for. It will also look things up, read pictures and files you attach, and remember what you tell it to keep. But writing is the job it was built and tested for.
If most of what you want help with is research, maths or code, a general assistant is the better fit. If most of it is the words that go out under your name, a writing model is.
How you use them
Gemini
You use Gemini the way you use most general assistants. You open it, type a request, and read the reply. Because it comes from Google, it also appears inside other Google products, so for many people it is simply there already.
Hemmingway
Hemmingway works in two ways.
- As a writer you talk to. Open the web app or the Mac, Windows or Android app, say who the message is for and what you want said, and it writes it. "Tell Sam I can't make Thursday, offer next week" is enough.
- As an inbox that drafts for you. In the apps, Inbox puts your mail and chats in one list: Gmail, Outlook, WhatsApp, Slack and Teams, and iMessage on the Mac. Open a message and Hemmingway drafts the reply. Nothing goes out until you send it. How that works.
Two parts of Hemmingway exist to make the draft sound like you. People holds who someone is to you (your boss, a client, your sister) and how you write to them, in the language you use with them. Memory holds the facts it should keep in mind, and you can read, change or delete every line. They never see any of it.
What we can say about Hemmingway-1's writing
We tested Hemmingway-1 against some of the largest models there are, and those results are on the model card. On eighty real email and message requests, in blind head-to-head matchups run in both orders, it came first on everyday writing (1026) and first on sounding like a person (1032). Those two tests, CommunicationBench and Human-Likeness, are our own benchmarks. On EQ-Bench 4, a public emotional-intelligence benchmark we ran with its official harness, it scored 1330, third, with Fable 5 higher at 1341.
None of that tells you how it compares with Gemini, because Gemini was not in the tests. It tells you how it compares with the models that were.
It also has limits. It is English-first. It can be wrong and still sound sure. For long fiction, larger story models beat it in our own StoryBench test. And it should not be trusted with medical, legal or money decisions.
Privacy
We never train our models on your chats, memory or People. On the web and on Android you can keep chats only on your device. Our privacy page lists everything we keep, in plain terms.
Google has its own settings and policies for Gemini. If privacy matters to your choice, read Google's current settings for your account rather than relying on a summary.
Open weights
Hemmingway-1 is open weights under Apache-2.0, including commercial use. You can download it from Hugging Face, run it on your own hardware, fine-tune it, or build a product on it. It has a 262,144-token context. If you would rather not run a 27B model yourself, the same model is behind an OpenAI-compatible API as hemmingway-27b: see the docs and the API platform.
The Gemini models you use in Google's assistant are not published as open weights. For many people that makes no difference. For a company that must keep text on its own machines, or a developer who wants to own the model they build on, it can decide the question.
Try HemmingwayDownload the app
How to test them yourself
Since we have no scores for Gemini, the fairest test is your own. It takes about fifteen minutes.
- Pick three messages you actually need to write this week. Mix them: one easy, one awkward, one to someone you know well.
- Write down the facts for each in a line or two: who it is for, what you want to say, any date, number or reason.
- Give both tools exactly the same words. For example: "Write the message I'll send my landlord. The boiler has been broken since Monday. I need it fixed by Friday. Only use these facts. Just the message, one version."
- Paste both replies into one document without labels, and come back to them later.
- For each, ask three questions. Would I send this with no edits? Did it add anything I did not give it, such as a day or a reason? Does it sound like me, or like an assistant?
Count how many you would send as they are. That number tells you more than any benchmark, because it is measured on your writing.
Which to use
- Research, questions, code, work inside Google's products: Gemini.
- Emails, texts and replies that have to sound like you: Hemmingway.
- A model you can download and run yourself: Hemmingway-1.
- Both: plenty of people use a general assistant for thinking and a writing tool for the words they send.
Common questions
Is Hemmingway better than Gemini for writing?
We cannot say, because we did not benchmark Gemini. What we can say is that Hemmingway-1 came first on everyday writing and on sounding like a person against the models we did test, in our own benchmarks. The quickest way to decide is to give both the same three messages and see which you would send unedited.
Is Hemmingway made by Google?
No. Hemmingway is its own lab and product, and Hemmingway-1 is our own 27B model, built on Qwen3.8-27B. It is also not the Hemingway Editor, and has nothing to do with Ernest Hemingway.
Does Hemmingway train on my chats?
No. We never train our models on your chats, memory or People. The privacy page lists what is kept. On the web and on Android you can also keep chats only on your device.
Can Hemmingway write replies to my Gmail?
Yes. In the Mac, Windows and Android apps you can connect Gmail in Inbox, and Hemmingway drafts each reply for you to read and send. It does not send anything in your name unless you have switched on Autopilot and the message fits your rules.
Is Hemmingway-1 open source?
Yes. The weights are on Hugging Face under Apache-2.0, which allows commercial use, and the code is on GitHub. You can also use it through an OpenAI-compatible API as hemmingway-27b.