A small lab's 27B model that beats every frontier model at writing like a person. #1 on human-likeness and everyday writing, third in the world on the public EQ-Bench 4, at a fraction of their size. Chat on the web, or let it draft your replies. The weights are open, Apache-2.0.

Hey Product Hunt, I'm Luka. We're a small lab, two of us plus a friend who does our marketing, and we spent this year training a 27B model to write like a person. It writes more like a person than any frontier model we put it up against, and it beats them on everyday writing too, the texts and the emails. EQ-Bench 4 is public and isn't ours, and on that one it comes third in the world. The charts are all on hemmingway.io/model/ if you want to see the boards. It's 27 billion parameters, which is small. The models it's beating are many times bigger, and the Qwen base we trained it from was nowhere near, so it's the training that did it and not the size. It isn't first at everything. On long fiction a couple of the big models are still ahead of it. We built it for the writing you actually have to do, which is answering people, and that's where it wins. One of our testers ran its texts through an AI detector next to Gemini, ChatGPT and GLM. Ours came back 0% both times, the others averaged over 60%. You can use it in the browser without installing anything. Tell it who a message is for and it keeps what it knows about that person and writes to them differently. On Mac and Android there's an Autopilot that reads your mail and chats and writes the replies, and it asks you before anything goes out. It takes voice and pictures too. Same model on the API. The weights are open as well, Apache-2.0, so you can run it yourself: huggingface.co/Altworld/Hemmingway-1 We don't train on your chats. That includes the ones you save to your account, your memory, and the people you write to. It's on the privacy page in those words. It isn't finished. It's slower than the big models and it still makes things up if you ask about something it doesn't know. About forty people have been testing it this week and we fix what they find. If it does something stupid, say so here and I'll answer. Luka

How did you train the model?

@stargiraffe We used a mix of SFT and Self distillation to bring out the most human-like behaviours of the model

@stargiraffe @luka234 Where did the SFT data come from: real human writing like emails and texts, or writing you collected or commissioned? And in the self-distillation step, how did you pick which outputs counted as more human-like? Was that human raters, a classifier, or the model judging itself?

@stargiraffe @koistya Most of the data was generated synthetically by us, and the big chunk of non-synthetic data was sourced from world-class Public-domain books and art. for the human-likeness we had human testers aswell as frontier OSS models judging it.

How did you train the model?

@stargiraffe We used a mix of SFT and Self distillation to bring out the most human-like behaviours of the model

Hemmingway AI

The best AI in the world at writing like a person

The best AI in the world at writing like a person