The best open-source LLM for writing

If you want a writing model you can run yourself and fine-tune, you need open weights and a licence that lets you use them. Hemmingway-1.5 and Hemmingway-1 have both: they are free for non-commercial use, commercial use is by agreement with us, and they were built for one job: writing that sounds like a person wrote it.

Hemmingway-1.5 is the newer one, released on 25 October 2026. 90% of its writing passes GPTZero & Pangram and the top AI detectors on the first try, across 1,848 texts in 13 kinds of writing. Real human writing passes the same detectors 86% of the time, and Hemmingway-1 passed 7%. How that was tested.

What they are

The code is on GitHub at github.com/lukeckprobierts/Hemmingway-1.

How Hemmingway-1 scores against open and closed models

The benchmark scores below are Hemmingway-1's. The most telling comparison for an open model is against the model it started from. Hemmingway-1 is Qwen3.8-27B trained further for writing, so the base model's scores show what the training did.

Against much larger models, on everyday writing it came first of all we tested, level with Fable 5.1 and ahead of Fable 5, GLM-5.3, Kimi K3, GPT-6 Astra, Grok 4.6 and DeepSeek V4 Pro. On the public EQ-Bench 4 it placed third. On long stories it sits level with Kimi K3, ahead of Qwen3.8 Max and DeepSeek V4 Pro, and behind Fable 5 Max and GLM-5.3. All of the scores are on Hemmingway-1's model card.

CommunicationBench, Human-Likeness and StoryBench are our own benchmarks. Every matchup was blind, run in both orders, and judged by a model that was not one of those being judged. EQ-Bench 4 is not ours.

Running it

With vLLM:

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

For Hemmingway-1, use HemmingwayAI/Hemmingway-1. Or load either with Transformers as a normal chat model, with AutoModelForCausalLM and AutoTokenizer, and build the prompt with the tokenizer's chat template. The guides to running Hemmingway-1.5 and running Hemmingway-1 have full examples.

Or skip the hardware

If you would rather not run a 27B model yourself, the same models are behind an OpenAI-compatible API as hemmingway-1.5, the default, and hemmingway-1. Most code that already talks to a model needs two lines changed. The API docs are here, and keys are on the API platform.

What they are not good at

They are English-first, and for very long fiction the largest story models are better.

Common questions

What is the best open-source LLM for writing?

We make open-weights models rather than open-source ones, and both are built for writing that sounds like a person. Hemmingway-1.5's writing passes GPTZero & Pangram and the top AI detectors 90% of the time. Hemmingway-1, the model it is built on, came first on our CommunicationBench of all the models we tested, open and closed, level with Fable 5.1. CommunicationBench is our own benchmark; the scores are on the Hemmingway-1.5 and Hemmingway-1 model cards.

Can I use Hemmingway-1.5 or Hemmingway-1 commercially?

Yes, with an agreement. Both are released under CC BY-NC 4.0, which is free for personal, research and other non-commercial use. For commercial use, write to luka@hemmingway.io; we are glad to work with anyone who asks. The weights are on Hugging Face at HemmingwayAI/Hemmingway-1.5 and HemmingwayAI/Hemmingway-1, and the code on GitHub.

How do I run Hemmingway-1.5 locally?

With vLLM, run vllm serve HemmingwayAI/Hemmingway-1.5 --max-model-len 262144, at temperature 1.0. Or load it with Transformers as a normal chat model, using AutoModelForCausalLM, AutoTokenizer and the tokenizer's chat template. The full guide.

What is Hemmingway-1.5's context length?

262,144 tokens, the same as Hemmingway-1. Both are 27B models.

Is there an API for Hemmingway-1.5?

Yes. It is behind an OpenAI-compatible API as hemmingway-1.5, the default, with Hemmingway-1 as hemmingway-1, so most existing code needs two lines changed. The API docs are here, and keys and prices are on the API platform.