diff --git a/recipies/gb200-fp4/1k8k/low-latency.yaml b/recipies/gb200-fp4/1k8k/low-latency.yaml
index a40ed9ba..b89d0982 100644
--- a/recipies/gb200-fp4/1k8k/low-latency.yaml
+++ b/recipies/gb200-fp4/1k8k/low-latency.yaml
@@ -2,7 +2,7 @@ name: "gb200-fp4-1p2d"
 
 model:
   path: "dsr1"
-  container: "lmsysorg/sglang:dev-cu13"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260121-1e309030"
   precision: "fp4"
 
 resources:
diff --git a/recipies/gb200-fp4/1k8k/max-tpt.yaml b/recipies/gb200-fp4/1k8k/max-tpt.yaml
index adbc3855..97f0f3f1 100644
--- a/recipies/gb200-fp4/1k8k/max-tpt.yaml
+++ b/recipies/gb200-fp4/1k8k/max-tpt.yaml
@@ -4,7 +4,7 @@ name: "gb200-fp4-max-tpt"
 
 model:
   path: "dsr1"
-  container: "lmsysorg/sglang:dev-cu13"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260121-1e309030"
   precision: "fp4"
 
 resources:
@@ -55,7 +55,6 @@ backend:
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: "1"
     SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
     SGLANG_MOE_NVFP4_DISPATCH: "1"
-    SGLANG_FLASHINFER_FP4_GEMM_BACKEND: "cutlass"
 
   sglang_config:
     prefill:
@@ -230,6 +229,7 @@ backend:
       enable-dp-lm-head: true
       prefill-round-robin-balance: true
       enable-dp-attention: true
+      fp4-gemm-backend: "flashinfer_cutlass"
 
       # Parallelism
       tp-size: 48
diff --git a/recipies/gb200-fp4/1k8k/mid-curve.yaml b/recipies/gb200-fp4/1k8k/mid-curve.yaml
index 358126cb..1e5771f2 100644
--- a/recipies/gb200-fp4/1k8k/mid-curve.yaml
+++ b/recipies/gb200-fp4/1k8k/mid-curve.yaml
@@ -5,7 +5,7 @@ name: "gb200-fp4-max-tpt-2"
 
 model:
   path: "dsr1"
-  container: "lmsysorg/sglang:dev-cu13"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260121-1e309030"
   precision: "fp4"
 
 resources:
@@ -56,7 +56,6 @@ backend:
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: "1"
     SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
     SGLANG_MOE_NVFP4_DISPATCH: "1"
-    SGLANG_FLASHINFER_FP4_GEMM_BACKEND: "cutlass"
 
   sglang_config:
     prefill:
@@ -230,6 +229,7 @@ backend:
       enable-dp-lm-head: true
       prefill-round-robin-balance: true
       enable-dp-attention: true
+      fp4-gemm-backend: "flashinfer_cutlass"
 
       # Parallelism
       tp-size: 32