In short: a model is a pile of numbers called weights. Quantization writes each of them with fewer digits. The file gets lighter, but the numbers are no longer quite the same.
Think of rounding. A price of 7.48 euros can be written as 7.5 and everyone knows what it means. That saves space. It also loses precision. How much depends on how you round and which numbers you round. That is why the same quantization is almost invisible on one model and noticeable on another. The only way to know is to run it and measure.
Example: NVFP4 on DeepSeek V4.1 Flash
On October 30, 2026, NVIDIA published DeepSeek V4.1 Flash quantized in NVFP4. According to NVIDIA, the checkpoint is about 400 GiB, while the MXFP4 original is about 475. The most instructive part is in the small print of the card: the expert weights went through the conversion without loss, while the Engram tables, the model's memory, were converted with loss. The card adds that lossless weights do not mean identical outputs at inference. So even a lossless conversion can still produce different output from the original. The check is on you: run the model with your own engine and compare on your own tasks.
The visual is generated code-art, no third-party images.