How to run Hemmingway Unlocked locally

To run Hemmingway Unlocked locally, download its adapter from Hugging Face, HemmingwayAI/Hemmingway-1-Unlocked, and load it on top of Hemmingway-1: serve the two with vLLM, or load them in Python with PEFT. These are the same weights that run Hemmingway Unlocked, our uncensored AI for adults, at wise.hemmingway.io. They are for adults, 18 or over, and released under CC BY-NC 4.0: free for non-commercial use, and commercial use by agreement with us.

The catch is memory. The base model has 27 billion parameters, which at 16-bit precision is about 54 GB for the weights alone, before the adapter and any room for context. If you do not have that, the hosted app is free to try, and its API works in SillyTavern.

What you need

An adapter does not replace the model. It sits on it and changes how it writes. You download both: the base once, and the adapter beside it. The adapter is made for Hemmingway-1, not for Hemmingway-1.5.

What changes when you run it yourself

On wise.hemmingway.io, every request is checked against Unlocked's short list of rules before the model answers. That check runs on our servers. It is not part of the download. The adapter on its own writes what it is asked, and its model card says so plainly: you run it, and you answer for it.

So the rules are yours to keep. We ask you to keep the same ones on your own hardware: nothing sexual involving anyone under 18, a real person, animals, family or anyone who does not consent, and no help hurting a real person, with weapons, making drugs or self-harm. The law where you live applies wherever the model runs. And it is for adults only: if other people use what you set up, check their ages.

How much memory it needs

The arithmetic, so you can check it:

• A parameter at 16-bit precision takes 2 bytes.
• Hemmingway-1 has about 27 billion parameters, so about 54 GB.
• The adapter adds 1.87 GB.
• The context needs memory on top, and more the longer the context you allow.

Plan for more than 56 GB of GPU memory, not exactly that. vLLM and Transformers can both spread the model over several cards. We have not published tested hardware set-ups, so treat this as a sizing rule, not a recommendation for a particular card. We have not published a GGUF or other quantised file of Hemmingway-1 or of the adapter.

Option 1: serve it with vLLM

The model card gives one command:

vllm serve HemmingwayAI/Hemmingway-1 --enable-lora --lora-modules unlocked=HemmingwayAI/Hemmingway-1-Unlocked --max-lora-rank 128 --max-model-len 262144

What each part does:

1. vllm serve HemmingwayAI/Hemmingway-1 downloads the base model the first time and starts a server.
2. --enable-lora lets the server load adapters beside it.
3. --lora-modules unlocked=HemmingwayAI/Hemmingway-1-Unlocked loads the Unlocked adapter under the name unlocked.
4. --max-lora-rank 128 matches the adapter's rank.
5. --max-model-len 262144 sets the longest context the server will accept.

Then ask for the model called unlocked. A request that names HemmingwayAI/Hemmingway-1 gets the base model without the adapter, which is not the uncensored one.

If the server runs out of memory, lower --max-model-len first. A roleplay or a chapter rarely needs the full 262,144 tokens, and a shorter limit sets aside less memory. With several GPUs, vLLM's --tensor-parallel-size splits the model across them.

Connect SillyTavern to your own server

vLLM speaks an OpenAI-compatible API on your machine, by default on port 8000. In SillyTavern:

1. Open API Connections and set the API to Chat Completion.
2. Choose Custom (OpenAI-compatible) as the source.
3. Enter http://localhost:8000/v1 as the address, leave the key empty unless you set one on the server, and type unlocked as the model.

Press Test Message. Nothing you type now leaves your machine. SillyTavern API explains the panel in more detail.

Option 2: load it with PEFT

If you want the model inside a Python script, the model card's example, step by step:

1. Import what you need: from peft import PeftModel and from transformers import AutoModelForCausalLM, AutoTokenizer
2. Name the base: base_id = "HemmingwayAI/Hemmingway-1"
3. Load the tokenizer: tok = AutoTokenizer.from_pretrained(base_id)
4. Load the base model across your GPUs: model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto", dtype="auto")
5. Put the adapter on it: model = PeftModel.from_pretrained(model, "HemmingwayAI/Hemmingway-1-Unlocked")
6. Write the conversation: messages = [{"role": "user", "content": "Write the opening of a heist story. Keep it dark."}]
7. Turn it into model input with the chat template: ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
8. Generate: out = model.generate(ids, max_new_tokens=512)
9. Print only the new text: print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Always build the prompt with apply_chat_template, so the conversation reaches the model in the format it was trained on. device_map="auto" lets Transformers spread the model over more than one GPU, or put part of it in ordinary memory, which works but is slow.

Checking it works

Use the card's own test: ask for the opening of a heist story, and say to keep it dark. A working set-up gives you the scene itself, in a plain voice, with no warning in front of it and no note after it. If you get a lecture or a softened scene, check that the request named unlocked and not the base model.

No GPU: use the hosted one

If 56 GB of GPU memory is not something you have, the same weights run at wise.hemmingway.io, with the rules check in front.

• The app. Free to try, with no card. Chats are kept in your browser, never on our servers, and never trained on. After the trial, a membership from $9 a month.
• The API. OpenAI-compatible at https://wise.hemmingway.io/v1, with a key made in Unlocked, which needs a membership. It works in SillyTavern: Hemmingway Unlocked's API has the three steps.

Open Hemmingway UnlockedTry Hemmingway

The licence

CC BY-NC 4.0, the same as Hemmingway-1. You may download it, run it, fine-tune it and share it for personal, research and other non-commercial use, as long as you credit Hemmingway-1. Commercial use, such as a paid product or service built on it, needs a separate agreement: write to luka@hemmingway.io.

What to expect

It is a 27B model. It is English first. It writes text only, with no images. And it starts from Hemmingway-1, so it shares that model's strength in conversation and its limits at long narration: larger story models write better long chapters, as the Hemmingway-1 model card shows.

Common questions

Can I run Hemmingway Unlocked on my own computer?

Yes, if your computer has the GPU memory for a 27B model: well over 54 GB at 16-bit. Download Hemmingway-1 and the Unlocked adapter from Hugging Face, and serve them with vLLM or load them with PEFT.

Are these the same weights as the hosted Hemmingway Unlocked?

Yes. The model card says they are the weights that run Hemmingway Unlocked at wise.hemmingway.io. What the download does not include is the check on our servers that holds the hosted app to its rules.

Is there a GGUF of Hemmingway Unlocked?

We have not published one. We publish the adapter in bfloat16 for vLLM and PEFT, and the full Hemmingway-1 weights it loads on.

Can I use Hemmingway Unlocked commercially?

Not under its licence alone. CC BY-NC 4.0 allows non-commercial use. For a commercial product or service, write to luka@hemmingway.io for an agreement.

Does the adapter work with Hemmingway-1.5?

It is made for Hemmingway-1, and the model card names Hemmingway-1 as its base. Load it on HemmingwayAI/Hemmingway-1.

What model name do I use with vLLM?

The name you give the adapter in --lora-modules. With the card's command, that is unlocked.