Your model, only smaller.

drinkme runs local AI models with losslessly compressed BF16 weights. They take up a quarter less memory, with every weight preserved bit for bit. That leaves more room for your model and can make generation faster, too.

Qwen3.8-27B drops from 55 GB to 39 GB, enough to fit a single 48 GB GPU, and can generate about 20% faster than it does uncompressed.

“What a curious feeling!” said Alice; “I must be shutting up like a telescope.”

Lewis Carroll, Alice’s Adventures in Wonderland (1865)

Serve it on your own machine and connect your favourite OpenAI- or Anthropic-compatible client.

Get started

Make room for a bigger model.

The room you save is yours to spend. Fit a model that wouldn’t load before, or keep the one you have and give it a longer context. drinkme has two compression profiles: sip, the default, and gulp, which is a little smaller but slower to decode.

“Now I’m the right size to get through the little door!”

Lewis Carroll, The Nursery “Alice” (1890)

Each model’s weights in GB: uncompressed BF16, and compressed with sip and with gulp
Model	uncompressed BF16BF16	sip	gulp

These are the weights alone. You’ll also need room for the KV cache and runtime.

Smaller, and often faster.

Where memory bandwidth is the limit, reading fewer bytes makes generation faster, too. Choose a GPU like yours to see how much.

GPU

Let’s get your model running.

You’ll need Linux with an NVIDIA or ROCm-supported AMD GPU, or a Mac with Apple silicon, plus Git and uv. Without a supported GPU, drinkme still runs on the CPU, just slowly.

On the first run, drinkme sets up PyTorch for your GPU and suggests a model that fits your memory. If the model has a published pack, drinkme downloads it, about a quarter smaller than the original, and checks it bit for bit against the original before serving. Otherwise it downloads the original and compresses it on your machine. Either way, it’s kept for next time.

Install & serve

Tangled

GitHub

git clone https://tangled.org/ninachaubal.com/drinkme &&cd drinkme &&uv sync &&uv run --no-sync drinkme serve

The first download can be tens of gigabytes. You can choose a different model at the prompt.

Already have a model in mind? Use uv run --no-sync drinkme serve --model Qwen3-8B.

Connect your client

In your client’s provider settings, choose the API it uses and set the base URL:

OpenAI · Chat Completions & Responses: http://127.0.0.1:3215/v1
Anthropic · Messages: http://127.0.0.1:3215
API key: Any non-empty value works locally, such as local.
Model: The name you served, such as Qwen3-8B. You can list the models you’re serving with curl http://127.0.0.1:3215/v1/models.

The Caterpillar, blue, on a mushroom, smoking a hookah, with Alice peeking over the edge. The Nursery Alice, hand-coloured by John Tenniel, 1890.

Setup guide Models Server options

The White Rabbit in a checked jacket, hurrying and looking at his pocket watch. John Tenniel, 1865.

Curiouser and curiouser? Measure your own.

Memory savings depend mostly on the model. Speed also depends on your hardware, which may not be on the chart yet. Measure it, then share the result so the next person with the same setup knows what to expect.

Measure your setup

Run this from the drinkme folder. It may download a second, larger model, so add --dry-run first to see the plan. Results stay local until you share them.
```
uv run --no-sync drinkme bench
```
Share your result optional

Replace you.example.com with your Bluesky or other atproto handle.
```
uv run --no-sync drinkme publish --handle you.example.com
```
Sign in when your browser opens. Your latest result is published to your account and joins the chart above. Then invite a friend to measure theirs.