diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index dc14bb2f5a..8f3974c785 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -5831,13 +5831,12 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp:
           ep: 1
           dp-attn: false
 
-# MiniMax-M3.1-Flash AgentX on GB200 with SGLang: day-zero image from the SGLang
-# cookbook (https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1),
-# manifest digest verified 2026-10-24.
+# MiniMax-M3.1-Flash AgentX on GB200 with SGLang: upstream arm64 CUDA 13.4
+# nightly, which carries the MiniMax-M3.1-Flash support of sgl-project/sglang#45925.
 # Seven-token DSpark (bundled draft); synthetic acceptance follows
 # infx/golden_al_distribution/minimaxm31flash_dspark.yaml.
 minimaxm31flash-fp4-gb200-sglang-agentic-dspark:
-  image: lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87
+  image: lmsysorg/sglang:nightly-cu134-20261026-5ef9901@sha256:64460b432679d4555b6579832454d8b9a21cc8591b2d16b340b94abb59de28e6
   model: MiniMaxAI/MiniMax-M3.1-Flash
   model-prefix: minimaxm31flash
   runner: cluster:gb200-nv
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 46c342ab82..ff3037619c 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -11616,3 +11616,10 @@
     - "Upstream vLLM: the DeepSeek-V4-Pro MI355X mono decode layer (vllm-project/vllm#61482); the CSA/HCA default chain is enqueued first at small batch sizes (#61378); the ROCm sparse prefill uses the shared prefill chunk plan (#58405); the C128 compressor skip is no longer baked into CUDA graphs (#59709)."
     - "Draft model data type: no change. The DSpark draft (mtp.0-2 of DeepSeek-V4-Pro-0813 at revision 72e1d3230f6c) runs as shipped with the checkpoint's FP8 quant config: FP8 E4M3 attention, shared-expert and main_proj weights with 128x128 UE8M0 block scales, MXFP4 routed experts, BF16 norms, router gate, confidence and Markov heads. Its KV cache inherits the target's fp8_ds_mla, and --monokernel skips draft layers, so the draft keeps vLLM's standard forward."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4526
+
+- config-keys:
+    - minimaxm31flash-fp4-gb200-sglang-agentic-dspark
+  description:
+    - "Update the SGLang image from dev-minimax-m31-flash to nightly-cu134-20261026-5ef9901, pinned by digest."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4559
+  scenario-type: [agentic-coding]