Nimbus-Labs/Nimbus-Orage-12B

Nimbus Orage 12B

Nimbus Orage 12B is an open-weight coding model for local inference. It accepts text, images, and audio and generates text. Image and audio inputs require the companion projector. Nimbus Labs fine-tuned Orage from Google's Gemma 4 12B IT QAT checkpoint.

Download

Artifact	Size	Use
Nimbus-Orage-12B-Q5_K_M.gguf	8.55 GB	Model weights; works on its own for text
mmproj-Nimbus-Orage-12B-BF16.gguf	175 MB	Add this file for image and audio input

Use a Gemma 4 Unified-compatible build of llama.cpp. To start the multimodal server:

llama-server -m Nimbus-Orage-12B-Q5_K_M.gguf --mmproj mmproj-Nimbus-Orage-12B-BF16.gguf

Image and audio input were checked with llama.cpp b10007 on two image color prompts and two short speech clips. These are functional checks, not quality benchmarks. Video was not tested. Runtime memory exceeds the download size and varies with context length. See SHA256SUMS to verify the files.

HumanEval results

Benchmark	Passed	Pass@1
HumanEval	154 / 164	93.9%
HumanEval+	149 / 164	90.9%

Evaluation setup: Q5_K_M on llama.cpp b10007, with thinking disabled, temperature 0, and a 16,384-token completion cap. One answer per task was scored against both test suites. No upstream checkpoint was evaluated under this setup.

Base model and license

Orage is derived from Google Gemma 4 12B IT QAT and distributed under Apache 2.0. See third-party notices for attribution. Review and test generated code before using it.