OpenRouter publishes technical guide on fp8 and fp4 quantization impacts

On November 11, 2026, OpenRouter published a technical explainer on large language model quantization, focusing on how different numeric precisions affect both model output quality and price. The article opens with a common scenario: when you call an open-weight model through an API, several providers may serve it at different precisions, such as 16-bit bf16, 8-bit fp8 or 4-bit fp4, and that precision can shape both the answers you get and what you pay. The piece goes on to describe LLM quantization as the process that creates these variants by storing a model's weights in fewer bits, so the model needs less memory and hardware, at some cost in accuracy. The trade-off is therefore framed as a straight exchange: lower memory footprint and cheaper infrastructure on one side, measurable accuracy loss on the other. Notably, the article discusses these differences in general terms, pointing out that precision can affect output quality and price without providing benchmarks, internal evaluations or price comparisons of its own. As a result, the piece reads as a primer on the mechanism of fp8 and fp4 quantization rather than a set of recommendations, leaving open which precision a given caller should pick. That is where the material stops: no follow-up measurements, customer guidance or provider-specific pricing tables are mentioned.