EuroLLM-22B-Instruct-2512-FP8-Dynamic

FP8 W8A8 quantized version of utter-project/EuroLLM-22B-Instruct-2512.

Accuracy Comparison

lm_eval --model vllm --model_args pretrained=EuroLLM-22B-Instruct-2512-FP8-Dynamic,add_bos_token=True --task gsm8k --num_fewshot 5 --batch_size auto --limit 250

Model	Filter	Exact Match	StdErr
EuroLLM-22B-Instruct-2512	flexible-extract	0.576	±0.0313
EuroLLM-22B-Instruct-2512	strict-match	0.536	±0.0316
EuroLLM-22B-Instruct-2512-FP8-Dynamic	flexible-extract	0.520	±0.0317
EuroLLM-22B-Instruct-2512-FP8-Dynamic	strict-match	0.504	±0.0317