diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml
index 3ee414f245..1884b3a997 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml
@@ -6,7 +6,7 @@ base:
     gpus_per_node: 8
   frontend:
     type: atomesh
-    container_image: rocm/atom-dev@sha256:f00a9dc588a380038fdad54cd57a956043cb316ddda54ebb1b62a24f8caa343b
+    container_image: rocm/atom-dev@sha256:b78300a6caf6c1fec3b1f27b4f7333326c9d34911bfc7ae229d13d2ebbc43884
     enable_multiple_frontends: false
   observability:
     enabled: false
@@ -74,8 +74,6 @@ base:
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'atom:'
       AIPERF_APPLY_CHAT_TEMPLATE: 'true'
 
-# One variant per point. Full graphs cover every verify batch up to 2x CONC
-# times the verify window (1 + draft tokens).
 override_tp8_c1:
   roles:
     agg:
@@ -99,7 +97,7 @@ override_tp8_dcp8_c14_lmcache:
     agg:
       args:
         decode-context-parallel-size: 8
-        num-speculative-tokens: 3
+        num-speculative-tokens: 4
         cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112]'
         extra-kv-connectors: &lmcache_offload
           - kv_connector: lmcache_offload
@@ -135,7 +133,7 @@ override_tp8_dcp8_c16_lmcache:
     agg:
       args:
         decode-context-parallel-size: 8
-        num-speculative-tokens: 3
+        num-speculative-tokens: 4
         cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128]'
         extra-kv-connectors: *lmcache_offload
       env:
@@ -155,15 +153,14 @@ override_tp8_dcp8_c48_lmcache:
     agg:
       args:
         decode-context-parallel-size: 8
-        method: null
-        draft-model: null
-        num-speculative-tokens: null
+        num-speculative-tokens: 4
         max-num-seqs: 96
         cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96]'
         extra-kv-connectors: *lmcache_offload
       env:
         <<: *lmcache_env
         LMCACHE_MAX_LOCAL_CPU_SIZE: '128'
+        ATOM_ENABLE_REPLAYSSM: '1'
         ATOM_PREFILL_DECODE_INTERVAL: '4'
         ATOM_PREFILL_DELAYER_MAX_QUEUE_MS: '5000'
   benchmark:
@@ -178,15 +175,14 @@ override_tp8_dcp8_c56_lmcache:
     agg:
       args:
         decode-context-parallel-size: 8
-        method: null
-        draft-model: null
-        num-speculative-tokens: null
+        num-speculative-tokens: 4
         max-num-seqs: 112
         cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112]'
         extra-kv-connectors: *lmcache_offload
       env:
         <<: *lmcache_env
         LMCACHE_MAX_LOCAL_CPU_SIZE: '192'
+        ATOM_ENABLE_REPLAYSSM: '1'
         AITER_REUSE_IDENTICAL_COMM_GROUPS: '1'
         ATOM_PREFILL_DECODE_INTERVAL: '4'
         ATOM_PREFILL_DELAYER_MAX_QUEUE_MS: '5000'
@@ -202,15 +198,14 @@ override_tp8_dcp8_c72_lmcache:
     agg:
       args:
         decode-context-parallel-size: 8
-        method: null
-        draft-model: null
-        num-speculative-tokens: null
+        num-speculative-tokens: 4
         max-num-seqs: 144
         cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144]'
         extra-kv-connectors: *lmcache_offload
       env:
         <<: *lmcache_env
         LMCACHE_MAX_LOCAL_CPU_SIZE: '192'
+        ATOM_ENABLE_REPLAYSSM: '1'
         AITER_REUSE_IDENTICAL_COMM_GROUPS: '1'
         ATOM_PREFILL_DECODE_INTERVAL: '4'
         ATOM_PREFILL_DELAYER_MAX_QUEUE_MS: '5000'
diff --git a/inferencex-e2e/configs/amd-master.yaml b/inferencex-e2e/configs/amd-master.yaml
index bc1f08f480..3270da4bfd 100644
--- a/inferencex-e2e/configs/amd-master.yaml
+++ b/inferencex-e2e/configs/amd-master.yaml
@@ -696,7 +696,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
 # speculative decoding. Acceptance is pinned to the committed golden curve in
 # infx/golden_al_distribution/kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml
 # at both draft lengths used here: 7 draft tokens -> AL 3.84 at concurrency 1 and 4,
-# 3 draft tokens -> AL 3.00 at concurrency 14 and 16.
+# 4 draft tokens -> AL 3.36 from concurrency 14 up.
 # Companion to kimik3-fp4-mi355x-vllm-agentic-mtp: same checkpoint, same runner.
 # TP8 only -- the 1.56 TB MXFP4 checkpoint is ~195 GB/GPU and TP4 cannot load,
 # which is also why there is no DP-attention arm.
@@ -706,7 +706,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
 #                             contexts, so dcp-size 8 shards the KV read across
 #                             all 8 GPUs and the LMCache DRAM tier backs the
 #                             paged KV.
-#   throughput  (48, 56, 72)  same shape, no draft model.
+#   throughput  (48, 56, 72)  same shape, with the draft model.
 # Two utilization blocks because the per-rank LMCache pool differs, and the
 # launcher divides the aggregate budget by TP:
 #   0.343 -> 1028 GB aggregate -> 128 GB/rank (conc 14 through 48).
@@ -714,7 +714,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
 #            wider in-flight window needs the deeper pool to keep the paged KV
 #            resident.
 kimik3-fp4-mi355x-atom-agentic-mtp:
-  image: rocm/atom-dev:nightly_202609251613
+  image: rocm/atom-dev:nightly_202610131505
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index ffb4097b55..cc0e253b3c 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9866,3 +9866,16 @@
   description:
     - "Upgrade agentx-harness from 1.0.9 to 1.0.10: pin SPAWN subagents to one client worker to prevent duplicate session state accumulating across workers. Metric collection, calculations, and exports are unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4119
+
+- config-keys:
+    - kimik3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Move the MI355X Kimi-K3 FP4 ATOM AgentX submission from image rocm/atom-dev:nightly_202609251613 (ATOM 0.1.7.dev36+g68e0df5e7) to rocm/atom-dev:nightly_202610131505 (ATOM 0.1.8.dev44+g661ef7661), and pin the recipe frontend to that tag's digest sha256:b78300a6caf6c1fec3b1f27b4f7333326c9d34911bfc7ae229d13d2ebbc43884."
+    - "Keep the published concurrency set [1, 4, 14, 16, 48, 56, 72]; no points are added or removed."
+    - "Widen the DSpark draft from 3 to 4 tokens at concurrency 14 and 16, which moves the golden acceptance length from 3.00 to 3.36."
+    - "Serve concurrency 48, 56 and 72 with the same 4-token DSpark draft and ReplaySSM (ATOM_ENABLE_REPLAYSSM=1) instead of no draft model. Concurrency 1 and 4 keep the 7-token draft at AL 3.84. max-num-seqs, the CUDA-graph ladders, dcp-size, LMCache sizing and the prefill-delayer settings are unchanged."
+    - "The new image carries the Kimi-K3 prefill gated RMSNorm pipeline (ROCm/ATOM#2458) and moves the LMCache wheel that the in-process lmcache_offload connector imports from 0.5.6.dev98 to 0.5.6.dev139 (ROCm/ATOM#2438)."
+    - "No change to the Inferact/Kimi-K3-DSpark draft's precision: the checkpoint is BF16 throughout, and at ATOM 661ef7661 online_quant_config still excludes every draft linear (layers.*, context_proj), so its weights and activations stay BF16 at every drafted point, including the newly drafted concurrency 48, 56 and 72, and it shares the target's FP8 KV cache."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4085