diff --git a/python/sglang/srt/model_loader/loader.py b/python/sglang/srt/model_loader/loader.py
index d233612bd..2cc109864 100644
--- a/python/sglang/srt/model_loader/loader.py
+++ b/python/sglang/srt/model_loader/loader.py
@@ -839,7 +839,13 @@ class DefaultModelLoader(BaseModelLoader):
         if is_nvfp4_online:
             # Scope exact FP4 quantization math to load-time conversion only;
             # restore the original environment before serving starts.
-            with temp_set_env(FLASHINFER_DISABLE_FP4_QUANT_FAST_MATH="1"):
+            with temp_set_env(
+                FLASHINFER_DISABLE_FP4_QUANT_FAST_MATH="1",
+                FLASHINFER_NVFP4_4OVER6="1",
+                FLASHINFER_NVFP4_4OVER6_E4M3_USE_256="0",
+                FLASHINFER_NVFP4_4OVER6_ERR_MODE="MSE",
+                FLASHINFER_NVFP4_4OVER6_ERR_USE_FAST_MATH="1",
+            ):
                 model.load_weights(weights)
             if target_device.type == "cuda":
                 torch.cuda.synchronize()