Maybe model is sensitive to quantization, by default ollama quantize it significantly.