Solari-GGUF
GGUF quantized version of Solari — a 500M parameter vision-language model fine-tuned for reduced hallucination on real-world images.
For full model details, training procedure, and benchmark analysis, see the Solari model card.
Model Details
- Base Model: Cubex11/Solari
- Original Base: SmolVLM2-500M-Video-Instruct
- Format: GGUF (for use with llama.cpp, Ollama, LM Studio, etc.)
- Fine-tuning Method: QLoRA + DPO on RLAIF-V (72K preference pairs)
- License: Apache-2.0
Benchmark Results
Solari improves on 7 out of 8 benchmarks vs the base model:
| Benchmark | Base | Solari | Change |
|---|---|---|---|
| POPE Overall | 82.67 | 85.08 | +2.41 |
| POPE Recall | 76.73 | 85.33 | +8.60 |
| AMBER Avg | 79.38 | 79.77 | +0.39 |
| A-OKVQA | 68.12 | 69.00 | +0.88 |
| MMStar | 38.33 | 39.60 | +1.27 |
| MMBench | 53.14 | 53.42 | +0.28 |
| RealWorldQA | 49.80 | 50.59 | +0.78 |
| HallusionBench | 27.58 | 28.14 | +0.56 |
| MME Perception | 1216 | 1119 | -97.7 |
Note: Benchmarks were evaluated on the full-precision model. GGUF quantization may cause minor performance differences.
Usage
With llama.cpp
# Download the GGUF file
huggingface-cli download Cubex11/Solari-GGUF --local-dir ./solari-gguf
# Run inference
./llama-cli -m ./solari-gguf/Solari.gguf -p "Describe this image" --image your_image.jpg
With Ollama
# Create a Modelfile
echo 'FROM ./Solari.gguf' > Modelfile
ollama create solari -f Modelfile
ollama run solari
Links
- Full-precision model: Cubex11/Solari
- Training dataset: RLAIF-V
- Base model: SmolVLM2-500M-Video-Instruct
Citation
@misc{solari2026,
title={Solari: Hallucination-Reduced Vision Language Model via QLoRA DPO on RLAIF-V},
author={Cubex11},
year={2026},
url={https://huggingface.co/Cubex11/Solari}
}