Testing 305 GGUF tags across 15 model lines revealed that 2.3% of official Ollama tags failed most HumanEval+ tasks. The study found that exact tensor reproducibility from original weights does not guarantee functional performance in low-bit quantizations.
HOW THIS AFFECTS YOU
●
builderYou must verify the functional integrity of quantized models rather than relying on registry tags.