Ask which open model writes best, and someone will send you a leaderboard. At first that looks like the answer.
Then you look at the top of it. The best open model on EQ-Bench’s creative writing board is Kimi K3. It has 2.8 trillion parameters. The smallest build anyone has published is 466 GB.
You are not running that on a laptop.
So I split the question in two. Which open models write well? And which of those can you run yourself?
Here is the short version. For everyday messages, Hemmingway-1 is the one model built for the job, on its makers’ word. For fiction at home, Muse-Glimmer-30B ranks highest among models that fit one GPU, and nobody built it for writing. The community’s fiction tunes of Gemma 4 have no independent scores at all. The biggest open models write best and are hosted-only for almost everyone.
One note before the list. I have not run these models. The scores come from EQ-Bench’s data files, which I checked on September 22, 2026. Everything else is the maker’s own claim from the model card, and I say so each time.
How did I choose?
Four rules. No model passes all of them, which is why the list has three groups.
- The licence allows commercial use without a revenue clause: Apache-2.0 or MIT.
- A 4-bit build is 20 GB or less, so it fits a 24 GB GPU or a 32 GB Mac.
- Either the maker says it is for writing, or an independent board has scored its writing.
- Ready-made GGUF or MLX builds exist.
Built for writing
Five models whose makers aim at prose. None of them is on an independent board yet.
(1) Hemmingway-1 is for messages and emails.
Altworld/Hemmingway-1 is a 27B fine-tune of Qwen3.8-27B under Apache-2.0. The makers trained it on “the writing people actually do every day” and to return only the message.
Its scores are the makers’ own. They say it loses on long stories. What is Hemmingway-1? goes through the claims.
A 4-bit build is 17.4 GB. You can try it in the makers’ app first.
(2) Gemma-4-31B-StyleTune changes one tensor.
Gryphe’s tune of Gemma 4 31B, Apache-2.0. Only the output layer was trained, on “100% narrative data.” The maker reports 60% fewer clichés per 100 words, and adds: “I might be biased towards my own data.”
The Q4_K_M build is 20.4 GB, just over my line. IQ4_XS is 17.9 GB.
(3) G4-MeroMero-v2-31B is for narrative roleplay.
zerofata’s tune, Apache-2.0, “designed for creative tasks, particularly narrative RP.”
The maker measured something useful. Asked for a story with a bare prompt, stock Gemma 4 reached for the same names and openings 99% of the time. The tune did it 66% of the time.
(4) Gemma4-Gutenberg-31B is for literary prose.
nbeerbower’s tune, Apache-2.0, trained on about 14,800 preference pairs from Project Gutenberg texts. The card describes “an active dispreference for ‘AI slop’ phrasing.” It has the weakest community signal of the five.
(5) Equinox-31B is for interactive fiction.
From Latitude, the makers of AI Dungeon, Apache-2.0. The training data is mostly second person, present tense. The card says it “occupies a middle ground by design.”
Not built for writing, but measured
Three general models with independent writing scores. On the creative writing board, higher Elo is better, and 133 models are ranked.
(6) Muse-Glimmer-30B ranks highest at this size.
Muse-Glimmer-30B is 22nd overall with 1790 Elo, the best of any model that fits one GPU. Its card credits Meta Superintelligence Lab and says it is “purpose-built for autonomous agentic tasks.” Writing is not mentioned.
Apache-2.0. The official 4-bit build is 16.8 GB, and Ollama has muse-glimmer:30b.
(7) Qwen3.8-27B is the base Hemmingway-1 started from.
Qwen3.8-27B is 33rd with 1668 Elo, and scores 53.3 on the longform board. Apache-2.0. A Q4_K_M build is 16.5 GB, and Ollama has qwen3.8:27b.
(8) Gemma 4 is the base of most fiction tunes.
Gemma 4 31B scores 1366 Elo and 56.5 on longform, second among models this size. The longform leader at this size is the older Qwen3.5-27B, at 59.0. Google’s card lists “poems, scripts, code, marketing copy, and email drafts” among its uses. Apache-2.0.
The 31B build is 18.3 GB and the 12B is 7.1 GB, which is the one to try on a 16 GB laptop. Ollama has gemma4:31b and gemma4:12b.
Open, and too big for your desk
(9) Kimi K3 and GLM-5.3 write best.
Kimi K3 is 4th overall with 2071 Elo. GLM-5.3 is 5th with 2064, and leads the open models on longform at 81.8. The three models above them are closed: GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.
Their smallest builds are 466 GB and 217 GB. Use them hosted, through their makers or OpenRouter. Both licences are MIT-style with extra terms for very large companies.
(10) DeepSeek V4 Pro is the one with a plain MIT licence.
DeepSeek V4 Pro scores 75.6 on longform. A 4-bit build is 850 GB.
What did I leave out?
- Artemis-31B from TheDrummer: the card declares no licence.
- Grok 4.6: no open weights. xAI has published only Grok 1 and 2.
- Mistral Small Creative: API-only, and deprecated since March 2026.
- Uncensored merges: they have the biggest download numbers, and their cards say “all use cases.” That is a different article.
How do you run them?
| Model | Easiest route | 4-bit size |
|---|---|---|
| Hemmingway-1 | llama.cpp or LM Studio | 17.4 GB |
| Muse-Glimmer-30B | ollama run muse-glimmer:30b | 16.8 GB |
| Qwen3.8-27B | ollama run qwen3.8:27b | 16.5 GB |
| Gemma 4 31B | ollama run gemma4:31b | 18.3 GB |
| Gemma 4 12B | ollama run gemma4:12b | 7.1 GB |
| The Gemma 4 tunes | GGUF from the model’s page, in LM Studio or llama.cpp | about 18 to 20 GB |
Leave 1 to 2 GB of memory above the file size. The makers of the fiction tunes suggest a temperature of 0.8 to 1.0 with a little min-p.
Which one should you try first?
For emails and messages, Hemmingway-1. It is the only one trained for them.
For fiction on your own machine, start with Muse-Glimmer-30B or Gemma 4 31B, which have scores you can check. Then try a tune against it with your own prompt. The tunes make strong claims and no one has measured them.
For livelier fiction, a base model is another route. Which Model Is Best at Creative Writing? The Worse One covers those.