Uluka Labs
Menu
AWS portal

Run big AI models on smaller, cheaper hardware.

Uluka Labs compresses AI models and builds the software to run them. A compressed model fits on a smaller GPU, so it costs less to run.

Try UlukaSee the results

Our first test: Gemma 4 26B

Original

51.6 GB

Google's 4-bit version

15.6 GB

Uluka Dango

8.7 GB5.9× smaller
Bars are drawn to scale against the memory of common GPUs. A model has to fit inside the GPU it runs on.
uluka-runtimeUluka Dango · cuda:0
0.0 tok/s0 tokensVRAM 8.7 / 16 GB
An example request to the runtime through its OpenAI-compatible API. Sample data, not a live model.

Four public benchmarks, out of 100

Google's 4-bit version 15.6 GB

70.4

AWQ INT4 17.2 GB

70.2

Uluka Dango 8.7 GB

69.9half the memory
Mean of MMLU, HellaSwag, GSM8K and IFEval, same questions and settings for every model. Uluka is level with both 4-bit builds while using about half their memory.

One endpoint, one month of GPU time

24 GB GPU what a 4-bit build needs

 

16 GB GPU what Uluka Dango fits on

  

That is up to   a year in potential savings, for one endpoint. Ten endpoints, ten times over.

AWS on-demand, Linux, US East, 730 hours a month. The same model on a smaller card is the whole saving — see the savings calculator.

What we do

We compress models to a custom architecture.

Our custom Uluka architecture reduces the memory a model needs while preserving its capability. In our first test, Gemma 4 26B went from 51.6 GB to 8.7 GB.

We build the software to run them.

Compressed models need their own software to run. Ours loads one onto a single GPU and serves it through an OpenAI-compatible API, so apps you already have keep working.

The Uluka app

A desktop app for asking questions about your own files, or getting a report from them. It can run its model on your own computer.

Download the app