OpenRouter test finds no detected accuracy difference across fp8, fp4 and fp16
OpenRouter analyzed 1,455 provider endpoints on October 20, 2026, finding that 865, or 59.5%, did not report a precision label. In a test of six Qwen3.8 27B endpoints on that day, it detected no accuracy difference among fp8, fp4 and fp16 across 240 scored tasks; all six scored between 98.3% and 99.2%. One fp4 endpoint ranked lowest on 40 additional multiplication tasks, but by a gap too small to rule out chance. Lower precision was generally cheaper, but not universally: 8-bit cost less than 16-bit for 14 of 24 compared models, while 4-bit cost less than 8-bit for 26 of 33. The quality comparison covered one model on one day, and OpenRouter could not determine which quantization methods the providers used.
In this limited test, precision labels did not predict accuracy, while endpoint prices varied substantially even at the same precision. OpenRouter recommends testing candidate endpoints on users’ own prompts and workloads before committing to a quantized option.
LLM Quantization: What fp8 and fp4 Do to Quality and Price
OpenRouter ·11/11/2026


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. The precision can affect both the answers and the price. LLM quantization creates these versions by storing a model’s weights in fewer bits, so the model needs less memory and hardware, at some cost in accuracy.
This post explains the formats and reports what we measured on OpenRouter on October 20, 2026, across 1,455 provider endpoints and six endpoints serving Qwen3.8 27B. On our test, we didn’t detect an accuracy difference between fp8, fp4 and fp16, so the precision label alone didn’t predict accuracy. Lower precision was usually cheaper, but not for every model.
#OpenRouter#Qwen3.8 27B#gpt-oss-120b#MiniMax M3#Kimi K2.6#Gemma 4 31B