Hemmingway-1 is a new AI model that only writes. At first that sounds like a step backwards.
It has 27 billion parameters. Claude and ChatGPT are many times larger. They write too, and they also code, search and read long documents.
So why would anyone want a small model that only writes?
Its makers’ answer: on short messages, it sounds more like a person than the big models do. They say it came first in their tests, ahead of Fable 5.1 and GPT-6 Astra.
I went through what that claim rests on.
Here is the short version. The model is real, open and cheap. Two of the wins are on tests its makers built themselves, and the judge was another model. On the one public benchmark they report third place, from their own run, and their table leaves out the model that leads the public board. On long stories it loses, and its makers say so.
One note before the list. I have not tested it against other models. Every Hemmingway-1 score below is the makers’ own. You can try it in your browser and judge the voice yourself.
What is Hemmingway-1?
A fine-tune of Qwen3.8-27B with open weights, sold as an app and an API at hemmingway.io.
(1) It is open.
The weights are on Hugging Face under Apache-2.0. Weights are the numbers a model learns in training. Publishing them means anyone can download the model, run it on their own machine, change it, or sell a product built on it. Claude and GPT stay behind their makers’ APIs.
The context is 262,144 tokens, and the longest reply is 32,768 tokens.
(2) It is small, and it was not trained from scratch.
27 billion parameters is small next to frontier models. The makers started from Qwen3.8-27B, an open model that already existed, and trained it further on writing. That is what a fine-tune is.
(3) It was trained on everyday writing.
The makers’ words: “messages, emails, the awkward note to a colleague, the thing you have been putting off.”
(4) It gives you the message and nothing else.
Ask most models for a text to your landlord and you get three options and a paragraph about the options. Hemmingway-1 is trained to return only the text.
(5) It thinks before it answers.
Thinking is on by default, at the highest level. You can turn it down.
Is it related to Hemingway Editor?
No. Hemingway Editor, with one m, highlights hard sentences in text you wrote. Hemmingway-1 writes the text.
Why two m’s? No reason. The makers: “It is just the name.”
Does it really beat Claude and ChatGPT?
On two of its makers’ tests, yes. On the public benchmark, third in their own run. On long stories, no.
| Test | Who built it | Hemmingway-1 | Next to it |
|---|---|---|---|
| Human-Likeness | Hemmingway | 1032, first | Fable 5.1 1006, GPT-6 Astra 964 |
| CommunicationBench | Hemmingway | 1026, first | Fable 5.1 1024, GPT-6 Astra 976 |
| EQ-Bench 4 | Public, run by Hemmingway | 1330, third in their run | Fable 5 1341, Kimi K3 1332 |
| StoryBench | Hemmingway | 1197 | Fable 5 Max 1277, GLM-5.3 1254 |
(6) Three of the four tests belong to the makers.
They say so themselves: “We built them, we ran them, and we are saying that up front.” Each test is eighty real requests, compared in blind pairs, in both orders.
(7) The judge is a model.
For Human-Likeness the judge was asked one question: “which of these two did a person write?” The judge was a different model from the ones being judged.
No panel of readers was involved. No AI detector was either, and the makers report no detector results.
(8) The lead on everyday messages is a tie.
1026 against 1024. The makers’ own write-up calls the two models “level within the margin.”
The clear gaps are 26 points on Human-Likeness, and 50 points over GPT-6 Astra.
(9) They ran the public benchmark themselves.
EQ-Bench is not theirs. But the run was, and thinking was on. The leaderboard turns reasoning off for most models. They call the comparison “close but not exact.”
(10) Their table leaves out the leader of the public board.
The public EQ-Bench 4 board lists Claude Opus 5 first, at 1385. It is not in the makers’ table, which starts with Fable 5 at 1341.
Hemmingway-1 is not on the public board at all. Next to the public numbers, 1330 would sit fourth, 55 points behind the leader. The makers describe it as “inside twelve points of the best model on the board.”
I do not know why Opus 5 is missing. Their write-up counts 27 models, and the public data file has 28. I checked the board on September 22, 2026; its data was generated on July 26.
None of this makes the results wrong. A 27B model level with frontier models on short messages is a real result if it holds up. Nobody outside the company has reproduced it yet.
Where does it fall short?
The makers list three limits. I would add one.
(11) It is English-first.
The makers say so. If you write in another language, this is not the model for it yet.
(12) It can be wrong and sound sure.
They warn against using it for medical, legal or financial decisions.
(13) Big models write better long fiction.
Fable 5 Max beat it by 80 points on StoryBench. The makers say it loses on “hostile storytelling and long story turns.”
(14) It has one voice, and it is not yours.
The voice is plain and sounds human. It does not know how you write until you give it your own paragraphs to follow.
Where can you use it?
Six places, from no setup to your own hardware.
(15) In your browser, free.
Our Hemmingway-1 demo needs no account. Replies come from the makers’ API. We add no system prompt.
(16) In the makers’ app.
It runs on the web, Mac, Windows and Android. You tell it who someone is, and it writes to them the way you would. It can also read your mail and chats and draft the replies. Pricing starts with a free trial, then $9, $29 and $79 a month.
(17) Through the API.
The model id is hemmingway-27b, and the API is OpenAI-compatible. It costs $0.24 per million input tokens and $0.90 per million output tokens. Use Hemmingway-1 in Obsidian, VS Code or Cursor shows where that endpoint plugs in.
(18) On your own computer.
The community has published small builds: GGUF files from bartowski for llama.cpp and LM Studio, and MLX builds for Apple Silicon. The usual 4-bit file is about 17 GB. How to run Hemmingway-1 locally has the commands.
(19) In StashBase, as a writing skill.
In StashBase, Hemmingway is a writing skill for the agent you already use, not a separate agent. Ask for a draft, and the agent writes it with Hemmingway-1 through StashBase’s hosted gateway. You sign in to StashBase and need no Hemmingway account. It is free for a limited time.
A chat box gives the model one line to go on. In StashBase it drafts next to your notes, sources and past writing. See how to choose an agent.
(20) As a humanizer, in your browser.
Sound Less Like AI, our free humanizer on this model, takes a draft you already have, up to 6,000 characters, and has the model write it again from its meaning, keeping every fact. No account. It shows no detector score, and we make no detector claim.
Is it a humanizer?
Not in the usual sense, and the difference matters.
A humanizer edits a draft that already exists. Humanizer, the widely used Claude skill, is a list of patterns taken from Wikipedia’s Signs of AI writing. When we tested it, it cut a 67-word launch announcement to 45 words and kept every fact. It kept the pitch too, because a list removes what is on the list. In a second test, the same agent with a plain editing request and no skill returned nearly the same text.
Hemmingway-1 works at the moment of writing. Given the brief behind that first test, it returned 32 words with nothing to delete. That was one run, and the reply still had two habits from Humanizer’s list. What it shows is where each tool does its work.
A humanizer is still the right tool in three cases.
(21) The draft is already yours.
Your facts and your argument are in it. A model writing from a prompt knows only what the prompt says.
(22) You want it to sound like you.
Humanizer can follow a few paragraphs of your own writing. Hemmingway-1 has one voice.
(23) You write in another language.
Hemmingway-1 is English-first.
AI humanizer or a better model goes through the full comparison. Humanizer and similar Claude skills covers setup and the alternatives.
Should you use it?
For the writing it was trained on, yes. The email you keep putting off. A reply that must be firm without being rude. A short piece that must not read like a press release.
Keep a general model for research, structure, code and long fiction. Hemmingway-1 vs Claude vs ChatGPT goes through the split.
And if what you want is your own voice, no model has it by default. Give it your writing to follow. How to Write with AI Without Losing Your Voice covers the method. If you would rather the text never left your machine, run it locally.
Frequently asked questions
Is Hemmingway-1 open source?
Its weights are open, under Apache-2.0, so you can download the model, change it and sell what you build on it. The training data and the training code are not published. That makes it open weights rather than open source in the sense the term has for software.
Is Hemmingway-1 free?
The weights are free to download and run on your own machine. Our browser demo and humanizer are free with no account. The makers' app starts with a free trial and then costs $9 to $79 a month, and their API is billed per token. Inside StashBase the Hemmingway writing skill is free for a limited time.