Open-source models for roleplay

People run open-source models for roleplay for three reasons: privacy, because a model on your own machine never sends your chats anywhere; control, because you choose the model, the instructions and the settings; and independence, because nobody can change the model under you or close your account. The usual set-up is a model served on your own hardware, with a separate chat app on top that speaks the OpenAI chat format.

Hemmingway-1 is one open model you can use this way. It is 27B, open weights under Apache-2.0, with a 262,144-token context. It scores 1330 on EQ-Bench 4, a public benchmark of emotional intelligence in conversation, which places it third in our comparison. It was built for talking to people and writing, though, not for roleplay, and in our own tests it loses on hostile storytelling and long story turns, where story models are better. This guide covers how to run it, and it works the same for most other open models.

Why run an open model for roleplay

• Privacy. When the model runs on your hardware, the conversation stays there. No service stores it, reads it or trains on it, because no service sees it.
• Control. You write the system prompt, set the sampling, and decide how much of the conversation the model sees. Nothing is hidden between you and the model.
• No account. A model on your own machine needs no sign-up, no key and no connection once the weights are downloaded.
• It stays the same. A hosted model can change from one week to the next. Weights you have downloaded do not.
• The licence. Many open models allow wide use. Hemmingway-1 is Apache-2.0, which includes commercial use.

The price of all this is hardware. A 27B model is large, and the section below says how large.

The two ways to use an open model

Both give you the same weights. The choice is mostly about whether you have the hardware and how much the privacy matters to you.

Option 1: run Hemmingway-1 on your own hardware

The weights are on Hugging Face at Altworld/Hemmingway-1. The simplest way to serve them is vLLM. The README gives one command:

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

That downloads the weights the first time, then starts a server with the model's full context. The server speaks an OpenAI-compatible API on your own machine, so a chat app that speaks the OpenAI format can point at it. Use Altworld/Hemmingway-1 as the model name.

Memory. At 16-bit precision, 27 billion parameters take roughly 54 GB for the weights alone, and the context needs more on top. That usually means one large data-centre GPU or several smaller ones. If the server runs out of memory, lower --max-model-len. A roleplay session rarely needs the full 262,144 tokens, and a smaller limit sets aside less memory. There is no official GGUF or quantised release of Hemmingway-1. The full guide to running it locally covers Transformers and multiple GPUs.

Option 2: use it through the API

If you do not have the hardware, the same model is behind an OpenAI-compatible API. A chat app that lets you set a custom OpenAI-format endpoint can use it with three settings:

• Base URL: https://hemmingway.io/v1
• Model: hemmingway-27b
• API key: one you make on the API platform. Keys start with hemmingway_live_.

This needs an account, and your chats go to our server to be answered. The API platform states that nothing you send and nothing the model writes is logged. For cost, see pricing. The API docs list every field.

We have not tested particular roleplay chat apps with the API, so we cannot promise any one of them works. If an app speaks OpenAI's Chat Completions format and lets you change the base URL, it should be able to connect.

Settings that help with roleplay

The API and a vLLM server both accept the usual sampling fields. These are the ones worth knowing for character chat:

• temperature sets how varied the writing is. Higher is looser and more surprising; lower is steadier.
• presence_penalty, frequency_penalty and repetition_penalty discourage the model from repeating itself, a common problem in long scenes.
• stop ends the reply when a given piece of text appears. If your chat app labels your turns with your character's name, putting that name here stops the model writing your lines.
• max_tokens caps the reply. On Hemmingway-1 this includes the model's thinking, so leave room for it.

Hemmingway-1 thinks before it answers, at its highest level unless you ask for less. On the API you can set reasoning_effort to low or medium, or turn thinking off with enable_thinking: false. The thinking arrives in reasoning_content, apart from the reply, so a chat app can keep it out of the story. The Thinking section of the docs has the details.

A worked example: setting up a character

Most chat apps have a system prompt or a character field. Whatever it is called, fill it in like this.

1. Give the character a name and a situation: You are Ines, a ferry captain on a northern lake, in her fifties, tired of tourists.
2. Say how she speaks: Short sentences. Dry. Never explains a joke.
3. Say what she wants and what she would never do: She wants to finish the season without trouble. She never lies about the weather.
4. Set the roles: I play a passenger. Write only Ines's words and actions. Never write mine.
5. Set the length: Under 120 words a reply. Stop when it is my turn.

Then start the scene with your character's first line. If Ines drifts out of character, say so and repeat one line from the set-up. In a long session, add a short summary of what has happened now and then, so the important facts stay in view.

Try HemmingwayDownload the app

How Hemmingway-1 compares for roleplay

EQ-Bench 4 is public; we ran it ourselves with its official harness. StoryBench is our own benchmark, blind and run in both orders. In our comparison, GLM-5.3 is the stronger story writer, and Kimi K3 is just ahead on EQ-Bench 4. Hemmingway-1 is strongest where a character has to read a person and answer well. It is weaker on hostile storytelling and long story turns, and for long, narration-heavy adventures a story model is the better pick. All the scores are on the model card. More on choosing a roleplay model.

It is English-first, and it can be wrong and still sound certain.

Common questions

What is the best open-source model for roleplay?

In our tests GLM-5.3 writes better long stories on StoryBench, and Kimi K3 is just ahead of Hemmingway-1 on EQ-Bench 4. Check each model's own page for whether its weights are published. Which suits you depends on whether your roleplay leans on long narration or on conversation between characters.

Can I run a roleplay model offline?

Yes, if you run an open model on your own hardware. Once the weights are downloaded, serving them with vLLM needs no account and no connection, and the conversation stays on your machine. For Hemmingway-1, plan for well over 54 GB of GPU memory.

Can I use Hemmingway-1 with a roleplay chat app?

If the app speaks OpenAI's Chat Completions format and lets you set a custom endpoint, point it at https://hemmingway.io/v1 with the model hemmingway-27b and a key from the API platform, or at your own vLLM server. We have not tested particular apps, so we cannot confirm any one of them.

Is Hemmingway-1 good for roleplay?

It is good at the conversational side, which is why it scores 1330 on EQ-Bench 4. It is weaker at hostile storytelling and long story turns, where story models do better in our tests. It was built for talking to people and writing, not as a roleplay model.