diff --git a/docs/docs/references/environment_variables.mdx b/docs/docs/references/environment_variables.mdx
index b4aef340c..fb84b5c76 100644
--- a/docs/docs/references/environment_variables.mdx
+++ b/docs/docs/references/environment_variables.mdx
@@ -354,7 +354,7 @@ SGLang supports various environment variables that can be used to configure its
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_ENABLE_MOE_DEFERRED_FINALIZE</code></td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Defer the MoE finalize step to overlap it with other work.</td>
-      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>true</code></td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_PATCH_TOKENIZER</code></td>
diff --git a/python/sglang/kernels/jit/csrc/moe/moe_finalize_fuse_shared.cu b/python/sglang/kernels/jit/csrc/moe/moe_finalize_fuse_shared.cu
index 8ae8072b8..2a96c8f2a 100644
--- a/python/sglang/kernels/jit/csrc/moe/moe_finalize_fuse_shared.cu
+++ b/python/sglang/kernels/jit/csrc/moe/moe_finalize_fuse_shared.cu
@@ -35,10 +35,10 @@
  * ``DeepseekV3MoE.forward``, and gives the downstream allreduce+rmsnorm
  * a clean PDL handoff.
  *
- * Expert-weight dtype is templated on ``TypeExpW`` so we support both the
- * bf16 and fp32 topk-weight paths (DSv3/K2.5 trtllm backends use fp32
- * because their ``_routing_logits_dtype = torch.float32``; other backends
- * use bf16).
+ * Expert-weight dtype is templated on ``TypeExpW`` so we accept both bf16
+ * and fp32 topk weights. The trtllm deferred-finalize path always feeds bf16
+ * (the trtllm-gen routing kernel emits bf16 for every routing method); fp32
+ * is kept for callers that produce topk weights in fp32.
  *
  * Expert-weight scale convention: in our target backends
  * (flashinfer trtllm nvfp4 + unquantized), ``apply_routed_scaling_factor_on_output``
diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index 99fa1f44b..eee06bdc2 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -1332,7 +1332,7 @@ class Envs:
     # Sglang Cache Dir
     SGLANG_CACHE_DIR = EnvStr(os.path.expanduser("~/.cache/sglang"))
     SGLANG_FLASHINFER_AUTOTUNE_CACHE = EnvBool(True)
-    SGLANG_ENABLE_MOE_DEFERRED_FINALIZE = EnvBool(False)
+    SGLANG_ENABLE_MOE_DEFERRED_FINALIZE = EnvBool(True)
 
     # Plugin system
     SGLANG_PLATFORM = EnvStr("")
diff --git a/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py b/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
index d02de06f2..f821d095c 100644
--- a/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
+++ b/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
@@ -1130,13 +1130,6 @@ def fused_experts_none_to_flashinfer_trtllm_fp4(
         result = trtllm_fp4_block_scale_moe(**moe_kwargs)
         if defer_finalize:
             gemm2_out, expert_weights, expanded_idx_to_permuted_idx = result[:3]
-            # FIXME(kpham-sgl): flashinfer sizes this buffer from routing_logits
-            # dtype (fp32 in DSv3 decode) but always writes bf16 weights into it.
-            # Reinterpret the live bf16 prefix. Fix upstream alloc to drop this,
-            # tracking in https://github.com/flashinfer-ai/flashinfer/issues/3595
-            if expert_weights.dtype == torch.float32:
-                n, k = expert_weights.shape
-                expert_weights = expert_weights.view(torch.bfloat16).view(-1, k)[:n]
             result = FlashInferTrtllmDeferredFinalizeOutput(
                 gemm2_out=gemm2_out,
                 expert_weights=expert_weights,