Nimbus Orage 12B
Nimbus Orage 12B is a local coding model fine-tuned from Google’s Gemma 4 12B IT QAT checkpoint. It generates text from text, images, or audio. For text, download just the model; for images or audio, add the matching projector.
Download and run
| File | Download | Use |
|---|---|---|
| Q5_K_M model | 8,547,267,520 bytes (8.55 GB) | Required |
| BF16 projector | 175,115,648 bytes (175 MB) | Add for images or audio |
Use a Gemma 4 Unified-compatible build of llama.cpp. The two files total 8.72 GB; runtime memory is additional and depends on context length. Verify both files with SHA256SUMS.
Text:
llama-server -m Nimbus-Orage-12B-Q5_K_M.gguf
Images and audio:
llama-server -m Nimbus-Orage-12B-Q5_K_M.gguf --mmproj mmproj-Nimbus-Orage-12B-BF16.gguf
On llama.cpp b10007, the release passed two image color checks and two short audio transcription checks. These confirm the input paths, not image or audio quality. Video was not tested.
Coding evaluation
| Benchmark | Passed | Pass@1 |
|---|---|---|
| HumanEval | 154 / 164 | 93.9% |
| HumanEval+ | 149 / 164 | 90.9% |
The same single answer per task was scored against both suites. The Q5_K_M file ran on llama.cpp b10007 with thinking disabled, temperature 0, and a 16,384-token completion cap. The upstream checkpoint was not evaluated under this setup. See the release manifest.
Base model and license
Orage is derived from Google Gemma 4 12B IT QAT. See the Apache 2.0 license and third-party notices. Review and test generated code before use.