diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 5abee0a3744b..552067c3db7e 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1284,20 +1284,6 @@ def _handle_model_specific_adjustments(self):
             # Set moe backend for DeepSeek
             if is_sm100_supported():
                 quant_method = get_quantization_config(hf_config)
-                quant_cfg = getattr(hf_config, "quantization_config", None) or {}
-                config_groups = quant_cfg.get("config_groups", {})
-                group0 = config_groups.get("group_0", {})
-                weights_cfg = group0.get("weights", {})
-                # this also apply to kimi k2.5
-                # since it follow the compressed tensor int4 recipe
-                # but not kimi k2 instruct or 0905 instruct.
-                is_kimi_k2_k25_thinking_int4 = (
-                    quant_method == "compressed-tensors"
-                    and weights_cfg.get("num_bits") == 4
-                    and weights_cfg.get("group_size") == 32
-                    and weights_cfg.get("strategy") == "group"
-                    and weights_cfg.get("type") == "int"
-                )
                 if self.quantization is None:
                     # Default DeepSeek V3/R1 native FP8 when not explicitly set,
                     # Because we need this condition for an assertion in
@@ -1312,20 +1298,12 @@ def _handle_model_specific_adjustments(self):
                 if (
                     self.moe_a2a_backend == "none"
                     and self.moe_runner_backend == "auto"
-                    and (
-                        self.quantization in ["fp8", "modelopt_fp8", "modelopt_fp4"]
-                        or is_kimi_k2_k25_thinking_int4
-                    )
+                    and self.quantization in ["fp8", "modelopt_fp8", "modelopt_fp4"]
                 ):
                     self.moe_runner_backend = "flashinfer_trtllm"
-                    if is_kimi_k2_k25_thinking_int4:
-                        logger.info(
-                            "Use flashinfer_trtllm as MoE runner backend on Blackwell for Kimi K2 / K2.5 thinking int4"
-                        )
-                    else:
-                        logger.info(
-                            "Use flashinfer_trtllm as MoE runner backend on sm100 for DeepseekV3ForCausalLM"
-                        )
+                    logger.info(
+                        "Use flashinfer_trtllm as MoE runner backend on sm100 for DeepseekV3ForCausalLM"
+                    )
             elif is_hip():
                 if not self.enable_dp_attention and self.nnodes == 1:
                     # TODO (Hubert): Put this back later