Hemmingway-1 vs Qwen3.8-27B

Hemmingway-1 is Qwen3.8-27B trained further for writing. The two models are the same size, 27 billion parameters, with the same 262,144-token context. What changed is how they write: in our tests Hemmingway-1 scores 1026 on everyday messages against the base model's 954, 1032 on sounding like a person against 952, and 1197 on stories against 693. That last gap is 504 points.

So if you want a 27B model for writing emails, messages and short stories, Hemmingway-1 is the better choice of the two by a wide margin. If you want a general model for other work, the base model is the general-purpose starting point. We tested both on writing only.

Same model underneath

Hemmingway-1 did not start from nothing. It is built on Qwen3.8-27B, a general-purpose model from Alibaba's Qwen family, and keeps its size and context length.

Because the size is the same, the hardware you need to run one is the hardware you need to run the other. Swapping one for the other in a vLLM setup is a change of model name.

The scores

CommunicationBench, Human-Likeness and StoryBench are our own benchmarks. We built them and ran them, and we say so on the model card. Every matchup is blind, run in both orders so position cannot sway the result, and judged by a model that is not one of those being judged. Scores are on an Elo scale: a higher score wins more of its head-to-head matchups, and the bigger the gap, the more lopsided those matchups are.

What changed in practice

Everyday messages

On CommunicationBench, eighty real email and message requests, the base model had the lowest score in the model card's table, 954. Hemmingway-1 placed first at 1026, level with Fable 5.1 and ahead of Fable 5, GLM-5.3, Kimi K3 and GPT-6 Astra. The fine-tuning took a 27B model from the bottom of that table to the top.

Sounding like a person

Human-Likeness asks the judge one question about each pair: which of these two did a person write? The base model scored 952. Hemmingway-1 scored 1032, first on the board and twenty-six points clear of the next model, Fable 5.1. This is the change the training was aimed at: writing that reads as if someone sat down and wrote it.

Stories

The largest change is on StoryBench. The base model scored 693. Hemmingway-1 scored 1197, level with Kimi K3 and ahead of Qwen3.8 Max at 1081 and DeepSeek V4 Pro at 1054. In our tests, then, the 27B model trained for writing beats Qwen3.8 Max, from its own family, on stories.

It is still not the best story writer. Fable 5 Max scored 1277 and GLM-5.3 1254, and for long fiction the big story models remain the better pick. More on creative writing.

What we can and cannot tell you

We can tell you what the two models are, what they score, and that Hemmingway-1 is the base model trained further for writing. How it was trained is beyond what this page covers, so we are not going to guess at it here.

We also only compared them on writing. We did not test the two on code, maths, or languages other than English, so we cannot tell you how they compare there. Hemmingway-1 is English-first. If your work is mostly not writing, test both on your own tasks before you choose.

A few other things apply to Hemmingway-1 whichever way you run it:

• It thinks before it answers, at its highest level unless you ask for less.
• It can be wrong and still sound certain. Do not use it to decide anything medical, legal or financial.
• For very long fiction, larger story models are better.

How to try both

If you already run Qwen3.8-27B with vLLM, you can run Hemmingway-1 the same way:

vllm serve Altworld/Hemmingway-1 --max-model-len 262144

Then give both models the same small set of your own requests. A good test set for writing:

1. Three emails or messages you actually needed to write last week, with the facts included.
2. One hard ask: a refusal, a request for money back, or bad news.
3. One short story prompt of about 150 words.
4. The same instruction on each: "Give me only the message, one version, no notes."

Read the answers aloud without knowing which model wrote which. Note which ones you would send as they are.

If you would rather not run a 27B model yourself, Hemmingway-1 is also behind an OpenAI-compatible API as hemmingway-27b, and in the Hemmingway apps. The API docs are here, and running it locally is covered here.

Try HemmingwayDownload the app

Which to use

• Emails, texts, replies and anything that has to sound like a person: Hemmingway-1.
• Short stories, letters and speeches: Hemmingway-1.
• A general 27B model for work other than writing: start from the base model, and test on your own tasks.
• Long fiction: neither is the strongest; a large story model will do better.

Common questions

Is Hemmingway-1 based on Qwen?

Yes. Hemmingway-1 is built on Qwen3.8-27B and trained further for writing. It has the same 27 billion parameters and the same 262,144-token context.

How much better is Hemmingway-1 than Qwen3.8-27B at writing?

In our own benchmarks it scores 1026 against 954 on everyday messages, 1032 against 952 on sounding like a person, and 1197 against 693 on stories, a gap of 504 points. These are our benchmarks, not public ones, and all of the scores are on the model card.

Does Hemmingway-1 need more hardware than Qwen3.8-27B?

No. They are the same size, so the hardware that runs one runs the other. With vLLM you change the model name to Altworld/Hemmingway-1.

Is Hemmingway-1 better than Qwen3.8-27B at coding?

We do not know. We tested the two on writing only. Hemmingway-1 is built for talking to people and writing, so for code, test both on your own tasks.

What licence is Hemmingway-1 under?

Apache-2.0, which allows commercial use. The weights are at huggingface.co/Altworld/Hemmingway-1 and the code at github.com/lukeckprobierts/Hemmingway-1.