Hemmingway-1.5: the writing model that passes AI detectors

Hemmingway-1.5 came out on 25 October 2026. It is our newest model, and the default in the Hemmingway app and the API. It has 27 billion parameters, it is built on Hemmingway-1, and its weights are open on Hugging Face under CC BY-NC 4.0.

The headline is one number. 90% of what Hemmingway-1.5 writes passes GPTZero and four of the top open AI detectors on the first try. Real human writing passes the same detectors 86% of the time. Hemmingway-1 passed 7%.

What changed from Hemmingway-1

Same size, same context, same licence. What changed is how it writes.

Hemmingway-1 wrote warm, plain messages and emails. AI detectors still caught nearly all of it. Hemmingway-1.5 is Hemmingway-1 trained further, and it writes the way people do across more kinds of writing: product descriptions, blog posts, reviews, forum comments, social posts, stories, speeches and emails. It is English-first.

It also rewrites. Paste in text from another AI, ask it to make it sound like a person wrote it, and it comes back in a person's voice.

Hemmingway-1 is still here. It is in the API as hemmingway-1 (the old id hemmingway-27b still works), in the app's model menu, and on Hugging Face. Its benchmark scores belong to it and stay on its own card: 1330 on EQ-Bench 4, third of 27, and first in our CommunicationBench with 1026.

How the detector test was run

• The texts. Hemmingway-1.5 wrote 1,848 texts across 13 kinds of writing.
• Every thinking level. It wrote them at all four thinking levels: off, low, medium and extra high.
• First tries only. Every text was written once, at temperature 1.0, and checked once. No picking, no retries.
• The detectors. GPTZero, and four of the top open AI-detector models: Fakespot from Mozilla, Vanguard, WasItAIGenerated and MELD. A text passes when their combined verdict says human with at least 90% confidence.
• The human baseline. 1,670 texts that people wrote, such as emails, essays, blog posts, reviews, stories and comments, went through the same detectors.
• The other models. Hemmingway-1.5 and the biggest models got the same 80 everyday asks, the asks from CommunicationBench, our own benchmark, and the same detectors.

The results

GPTZero passed it 90% of the time at every thinking level: off, low, medium and extra high alike. Written from scratch, 92% of its texts passed. Given AI text to rewrite, 89% of its rewrites passed.

By kind of writing, every kind passed more often than real human writing did:

On the same 80 everyday asks, the other models were caught most of the time:

Passing a detector is not the same as writing well, so we checked that too. Our quality judge scores its writing 4.2 out of 5 with thinking on, and 4.5 out of 5 for doing what it was asked. With thinking on, 64% of its first tries pass the detectors, write well and do what was asked, all at once.

Here is one first try from the test, as it was written, with thinking on extra high. The ask was a product description for a handmade ceramic coffee mug:

Handmade from Colorado red clay. Glazed in one of our own mixed colors, produced in small batches for true color variations. Please enjoy the little flaws. 12 oz (500ml). Dishwasher and microwave safe.

The quality judge and CommunicationBench are ours, and we say so wherever we quote them. Every result and more real outputs are on the model card.

Download it

The weights are at huggingface.co/HemmingwayAI/Hemmingway-1.5 under CC BY-NC 4.0: free for personal, research and other non-commercial use. Commercial use is by agreement: write to luka@hemmingway.io.

With vLLM, one command serves it:

vllm serve HemmingwayAI/Hemmingway-1.5 --max-model-len 262144

Use temperature 1.0. A lower temperature makes its writing easier to detect. The guide to running it yourself has the Transformers example, the thinking settings and the memory a 27B model needs.

Use it in the API

The API is OpenAI-compatible, at https://hemmingway.io/v1. The model id is hemmingway-1.5, and it is the default: a request that names no model gets it. Code that already uses OpenAI's SDKs needs only the base URL, a key and the model name.

It runs at temperature 1.0 when you send none, as it was tested. It reads pictures, supports function calling in OpenAI's format, and writes up to 32,768 tokens a reply, thinking included. Keys and prices are on the API platform, and every field is in the docs.

Use it in the app

Hemmingway-1.5 is the default in the Hemmingway app on Mac, Windows, Android and the web. It is free to try. In the apps it also drafts replies to your mail and chats in the way you write to each person, and on Mac and Android, Autopilot can answer the easy ones by rules you set. Autopilot is off until you turn it on.

Try HemmingwayDownload the app

Common questions

What is Hemmingway-1.5?

Hemmingway-1.5 is a 27B writing model from Hemmingway AI (Pty) Ltd, released on 25 October 2026. It is built on Hemmingway-1, has open weights under CC BY-NC 4.0, and 90% of its writing passes GPTZero and the top AI detectors on the first try.

When was Hemmingway-1.5 released?

On 25 October 2026. It became the default model in the Hemmingway app and the API that day.

Does Hemmingway-1.5 pass GPTZero?

In our test it passed GPTZero 90% of the time, at every thinking level. Across GPTZero and four open AI detectors together it also passed 90%, over 1,848 texts, each a first try.

What is the API model name for Hemmingway-1.5?

hemmingway-1.5, at https://hemmingway.io/v1. It is the default model, so a request with no model gets it. Hemmingway-1 is hemmingway-1.

Is Hemmingway-1.5 open source?

It has open weights, not open source. The weights are on Hugging Face at HemmingwayAI/Hemmingway-1.5 under CC BY-NC 4.0, free for non-commercial use. Commercial use is by agreement with us.

Is Hemmingway-1 still available?

Yes. Hemmingway-1 is in the API as hemmingway-1, in the app's model menu, and its weights are on Hugging Face at HemmingwayAI/Hemmingway-1.