Is the Model Dumber or the Provider Different? LLM Quantization (fp8, fp4)
The same model can swing between good and bad because inference providers quantize differently. We count fp8, fp4 and int4 across 137 OpenRouter endpoints and show how to pin precision.