diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh
index 9e5f8fc8d6..7f6f50f070 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh
@@ -116,7 +116,20 @@ else
     SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\", \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
 fi
 
-MAX_NUM_SEQS=$((2 * CONC))
+# Eval-only runs replay GSM8K through lm_eval at CONC parallel requests
+# (benchmark_lib run_eval: num_concurrent = EVAL_CONCURRENT_REQUESTS or CONC).
+# At CONC 1-3 that is 1319 sequential-ish requests at ~7 s each on H100 and
+# the eval overran the launcher's 180-minute allocation ("CANCELLED ... DUE TO
+# TIME LIMIT" at 96-98%, run 34174431989), while the recipe's previously
+# merged sweep only ever ran the eval at c8 (195 min). Accuracy does not depend
+# on client concurrency, so floor it at 8 for eval-only runs and size
+# --max-num-seqs to admit that many; throughput runs keep 2*CONC.
+EVAL_CONC=$CONC
+if [ "$EVAL_ONLY" = "true" ] && [ "$CONC" -lt 8 ]; then
+    EVAL_CONC=8
+    export EVAL_CONCURRENT_REQUESTS="${EVAL_CONCURRENT_REQUESTS:-$EVAL_CONC}"
+fi
+MAX_NUM_SEQS=$((2 * EVAL_CONC))
 # MTP verifies four tokens per sequence, so CUDA graph capture is token-sized.
 MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * TOKENS_PER_SEQ))
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 0cbe0e8c8f..333b523a25 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7514,7 +7514,7 @@ qwen3.5-fp4-b200-trt-mtp:
       - { tp: 8, ep: 8, dp-attn: true, spec-decoding: "mtp", conc-list: [128, 256, 1024] }
 
 minimaxm3-fp8-h100-vllm-agentic-mtp:
-  image: vllm/vllm-openai:v0.27.1
+  image: vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
   runner: cluster:h100-dgxc
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 4fcd229bfb..a54962b272 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6958,3 +6958,12 @@
     - "Expand the TP8 and TP4/EP1 sweep coverage through concurrency 640."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2866
   
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update vLLM image from vllm/vllm-openai:v0.27.1 (v0.27.1 release) to vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 upstream nightly, digest sha256:254eebf919e8b7b0d530d97fccc606c36f380ff6724d64bc190951bec1aee838, tag commit vllm-project/vllm@d9105ea8; Docker Hub last pushed 2026-09-07T06:16:01Z), the same tag the B200 MiniMax-M3 AgentX recipe moved to in #2860 and the ROCm counterpart of the MI325X/MI300X bumps in #2872/#2873. benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh is unchanged: TRITON_ATTN attention, fp8 KV, EAGLE3 with the Inferact MiniMax-M3 EAGLE3-GQA draft pinned to FLASH_ATTN and the committed golden synthetic acceptance length 2.78, Mooncake 0.3.11.post1 DRAM offload on the host-tier arm; the resident TP8 c1-c5 and Mooncake DRAM offload c6/c8 grid is unchanged."
+    - "Floor the eval-only client concurrency at 8 (EVAL_CONCURRENT_REQUESTS) and size --max-num-seqs to admit it: the eval matrix now runs GSM8K at every sweep concurrency, and at c1-c3 the H100 eval ran 1319 requests nearly serially at ~7 s each and overran the launcher's 180-minute Slurm allocation (run 34174431989: CANCELLED DUE TO TIME LIMIT at 96-98%; c4/c5 finished in 119-124 min). The previously merged sweep only ran the eval at c8 (195 min on v0.27.1), so this is not an image regression. Eval accuracy is independent of client concurrency; throughput runs keep max-num-seqs 2*CONC."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2874