Nimbus Labs · Available now

Nimbus Orage.

Your code.
Your machine.

Orage 12B is an open-weight coding model you can run locally. It accepts text on its own; add the companion projector to work with images and audio.

8.55 GB GGUF

175 MB projector

93.9% HumanEval

Direct mode · Q5_K_M · Pass@1

Coding results.

Orage passed 154 of 164 HumanEval tasks (93.9%) and 149 of 164 HumanEval+ tasks (90.9%). EvalPlus scored the same single answer per task against both test suites.

Compare released Nimbus coding scores

Local setup

Two files.
Three input types.

Run Orage with a Gemma 4 Unified-compatible build of llama.cpp. Download just the model for text, or both files for image and audio input.

Text weights
8.55 GB Q5_K_M GGUF. Runtime memory also depends on context length and the inference app.
Image and audio projector
175 MB extra, for 8.72 GB total on disk. Both input types passed functional checks in llama.cpp b10007; video was not tested.
Base model
Orage was fine-tuned from Google Gemma 4 12B IT QAT. Its coding scores are not a matched comparison with the untouched checkpoint.

Your code. Your machine.

Run Orage locally.

The model card includes setup instructions, checksums, evaluation settings, and attribution.

Open the model card