Ollada - run open diffusion language models on your own computer.
https://opensourcerulesz.github.io/ollada
https://github.com/OpenSourceRulesz/ollada
Most chatbots write word by word. A diffusion language model writes a whole answer at once, then fixes it over a few passes.
The problem is those models are scattered across Hugging Face, each with different, incompatible code, and nobody tells you if one will actually run on your machine. Ollada fixes that: one shelf, short names, one command, and every model is only listed after we've actually run it and measured it.
Try it, no GPU needed:
pip install "ollada[cpu]"
ollada pull sdar:1.7b
ollada run sdar:1.7b
If you get stuck, step-by-step setup instructions are on the site.
Five models: diffcoder:8b for coding, dream:7b, sdar:4b, sdar:1.7b, and nemotron:3b from NVIDIA. On a 16GB RTX 5060 Ti we measured DiffCoder 13.65 GB, Dream 9.13 GB, SDAR-4B 8.34 GB. On a GPU, start with sdar:4b. The SDAR models and Nemotron also run on a laptop CPU, slower but no GPU required.
Honest limits: no token streaming yet, CPU runs are slow on long answers, Dream is eh
Free, local, $0. Weights come from Hugging Face
I’m open to any conversations !
Top comments (0)