diff --git a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
index faf2232726..1d9bff1320 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
@@ -240,9 +240,9 @@ VLLM_CMD=(
     --reasoning-parser kimi_k3
     --tool-call-parser kimi_k3
     --enable-auto-tool-choice
-    # FP8 KV cache requires the prefill query quantization flag. MLA prefill
-    # runs on TRTLLM_RAGGED per the upstream Blackwell override.
-    --attention-config '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+    # FP8 KV cache requires prefill query quantization. Use FlashInfer MLA
+    # prefill, matching the current published upstream Kimi-K3 recipe.
+    --attention-config '{"mla_prefill_backend":"flashinfer","use_prefill_query_quantization":true}'
     --speculative-config "$SPEC_CONFIG"
     --compilation-config "$COMPILATION_CONFIG"
     --disable-uvicorn-access-log
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f916be18dc..a376a9a16b 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1521,7 +1521,7 @@ kimik3-fp4-b300-vllm-agentic-dspark:
   # its native window is 32k YaRN-stretched to 1M, and the same image/draft/config
   # reaches AL 4.18 on short-context work. These throughput numbers measure the
   # system at a prescribed acceptance, not the draft's fitness at 100k+ context.
-  image: vllm/vllm-openai:kimi-k3
+  image: vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:b300-nv
@@ -1533,17 +1533,15 @@ kimik3-fp4-b300-vllm-agentic-dspark:
     agentic-coding:
     - dram-utilization: 0.63
       search-space:
-      # Mirrors the non-MTP entry's KV arms so the spec-decoding delta is
-      # readable at equal concurrency, but stops at conc 16 rather than 24: the
-      # non-MTP bring-up sweep (run 30326393603) showed the GPU-resident arm
-      # already thrashing at conc >= 16 (prefix cache hit 2.7%, TTFT p50 86-191s)
-      # because GPU KV holds only ~3.1 max-length requests, so conc 24 costs a
-      # full job per arm to re-measure that regime. conc 16 still exercises the
-      # DRAM tier meaningfully (62% external prefix cache hit rate there).
-      # TP8-only for the same memory reason: a ~1.5 TB MXFP4 checkpoint does not
-      # fit below 8 GPUs.
+      # Keep equal resident/offload points through conc 8. The digest-pinned
+      # bring-up sweep (run 31404943911) confirmed that resident conc 16
+      # remained in warmup after 1,770 seconds (96/177 requests), while conc 8
+      # entered profiling. The offload conc 16 arm completed warmup in 804
+      # seconds and exercised the external KV tier, so retain it as the
+      # capacity endpoint.
+      # TP8-only: a ~1.5 TB MXFP4 checkpoint does not fit below 8 GPUs.
       # TP8 GPU-resident
-      - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 16] }
+      - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8] }
       # TP8 SimpleCPUOffload (host DRAM)
       - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [1, 2, 4, 8, 16] }
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 41e7238f8d..7ae8479441 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5784,3 +5784,13 @@
     - "Update the B300 DSV4 AgentX MTP image and SimpleCPU sweep through c576."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2536
   
+
+- config-keys:
+    - kimik3-fp4-b300-vllm-agentic-dspark
+  description:
+    - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
+    - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
+    - "Match the current published upstream Kimi K3 FP8-KV profile by using flashinfer MLA prefill with prefill query quantization"
+    - "Keep DSpark K=2 probabilistic throughput on synthetic golden AL 2.51 while the generated EVAL_ONLY row uses real block verification"
+    - "Cap the GPU-resident search at concurrency 8 after concurrency 16 failed to complete deterministic warmup; retain concurrency 16 for the DRAM-offload capacity tier"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476