diff --git a/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh
new file mode 100755
index 000000000..de36b5964
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh
@@ -0,0 +1,198 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# AgentX trace replay for Qwen3.8-Flash-Next NVFP4 on B200 with SGLang
+# native NEXTN MTP. Day-zero recipe; SGLang is the plan-of-record engine for
+# this model (MODELS.md). Throughput uses the golden synthetic AL; evals retain
+# real target-model verification.
+#
+# The checkpoint is RadixArk/Qwen3.8-Flash-Next-NVFP4 (126 GiB,
+# quantization_config.quant_method = modelopt), so --quantization modelopt_fp4
+# matches the same flag the Qwen3.5 NVFP4 sibling uses. The model ships native
+# MTP modules (kept unquantized by the checkpoint's ignore list), so NEXTN
+# needs no external drafter.
+#
+# TP1: the cookbook's verified single-node command for this model is --tp 1 on
+# both Blackwell parts. 126 GiB of NVFP4 weights fit on one B200, so the
+# model is not sharded and every rank-crossing collective disappears.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+# Use the lightweight GSM8K eval instead of the AgentX SWE-bench default.
+export EVAL_FRAMEWORK="lm-eval"
+
+check_env_vars \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING \
+    TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+
+SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+nvidia-smi
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126_256k
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+CACHE_ARGS=()
+if require_agentic_kv_offload_backend hicache; then
+    REQUESTED_HICACHE_TOTAL_GB="${HICACHE_TOTAL_CPU_DRAM_GB:-$TOTAL_CPU_DRAM_GB}"
+    if [ "$REQUESTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: requested HiCache pool ${REQUESTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    TOTAL_CPU_DRAM_GB="$REQUESTED_HICACHE_TOTAL_GB"
+    # SGLang applies --hicache-size independently to Qwen's target KV and
+    # Mamba pools. Native NEXTN also creates a draft KV pool with the same
+    # slot count; its one attention layer adds 1/15 of the target KV bytes.
+    # Reserve 1 GB/rank for page alignment and enforce H * 31/15 per rank.
+    HICACHE_ALIGNMENT_RESERVE_GB=$TP
+    HICACHE_USABLE_TOTAL_GB=$((TOTAL_CPU_DRAM_GB - HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$HICACHE_USABLE_TOTAL_GB" -lt 1 ]; then
+        echo "Error: insufficient DRAM after HiCache alignment reserve" >&2
+        exit 1
+    fi
+    MAX_HICACHE_SIZE_GB=$((HICACHE_USABLE_TOTAL_GB * 15 / TP / 31))
+    HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
+    if [ "$HICACHE_SIZE_GB" -lt 1 ] || [ "$HICACHE_SIZE_GB" -gt "$MAX_HICACHE_SIZE_GB" ]; then
+        echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB outside 1..$MAX_HICACHE_SIZE_GB" >&2
+        exit 1
+    fi
+    PROJECTED_HICACHE_TOTAL_GB=$(((HICACHE_SIZE_GB * TP * 31 + 14) / 15 + HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$PROJECTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: projected HiCache use ${PROJECTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB target + Mamba + 1/15 draft per rank across TP=${TP}; projected node total ${PROJECTED_HICACHE_TOTAL_GB} GB <= ${TOTAL_CPU_DRAM_GB} GB"
+    CACHE_ARGS=(
+        --page-size 64
+        --enable-hierarchical-cache
+        --hicache-size "$HICACHE_SIZE_GB"
+        --hicache-io-backend kernel
+        --hicache-mem-layout page_first
+        --hicache-write-policy write_through_selective
+    )
+fi
+
+PARALLEL_ARGS=(
+    --tp "$TP"
+    --dp 1
+    --ep-size "$EP_SIZE"
+)
+
+# TP4 needs parallel tokenization to keep 256k AgentX warmups below the client
+# request timeout. Keep TP2 on SGLang's single-worker default: multi-tokenizer
+# startup races with the TP2 HiCache shared-memory initialization path.
+TOKENIZER_ARGS=()
+if [ "$TP" -ge 4 ]; then
+    TOKENIZER_ARGS=(--tokenizer-worker-num 6)
+fi
+
+# AgentX concurrency counts live session trees rather than individual HTTP
+# requests. Leave room for subagent fan-out and avoid spending HBM on graphs
+# above the batch sizes that remain useful for this long-context workload.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS="$CONC"
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+
+export TORCH_CUDA_ARCH_LIST="10.0"
+export PYTHONNOUSERSITE=1
+export NCCL_NVLS_ENABLE=1
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+# Keep server-side connections alive beyond AIPerf's 300-second client pool
+# timeout so bursty AgentX trajectories cannot reuse a closing idle socket.
+export SGLANG_TIMEOUT_KEEP_ALIVE=1800
+
+if [ "${EVAL_ONLY:-false}" != "true" ]; then
+    # golden_al_distribution/qwen3.8next_mtp.yaml:
+    # qwen3.8-flash-next-fp8.thinking_on[3] = 2.32.
+    # --speculative-num-steps 3 with 4 draft tokens is 3 speculative tokens
+    # per verification step, i.e. the MTP=3 cell. AgentX replays run with
+    # thinking on, so the thinking_on row is the right one.
+    export SGLANG_SIMULATE_ACC_LEN=2.32
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    "${PARALLEL_ARGS[@]}"
+    # Verified flags from the SGLang cookbook playground for this model on
+    # B200 / NVFP4 / single node. Quantization is read from the
+    # checkpoint, so no --quantization flag; the hybrid GDN linear-attention
+    # layers take their own backends rather than --attention-backend.
+    --linear-attn-prefill-backend flashinfer
+    --linear-attn-decode-backend flashinfer
+    # bfloat16 is mandatory on Blackwell: SGLang rejects the launch outright
+    # with "--linear-attn-decode-backend flashinfer on SM100+ requires
+    # --mamba-ssm-dtype bfloat16". Hopper wants the opposite -- flashinfer's
+    # gated_delta_rule_mtp verify kernel asserts a float32 state there -- so
+    # the H200 arm sets float32 and this one must not follow it.
+    --mamba-ssm-dtype bfloat16
+    --speculative-algorithm NEXTN
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+    --reasoning-parser auto
+    # NEXTN silently resets --max-running-requests to 48 when it is unset, so
+    # this must stay explicit and sized to the AgentX concurrency.
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --mem-fraction-static 0.80
+    --stream-interval 50
+    --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
+    "${TOKENIZER_ARGS[@]}"
+    --tokenizer-path "$MODEL"
+    --enable-metrics
+    --enable-cache-report
+    "${CACHE_ARGS[@]}"
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+capture_cache_metrics() {
+    {
+        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
+        curl -fsS "http://localhost:$PORT/metrics" 2>/dev/null \
+            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
+            || true
+        echo "============================================================"
+    } >> "$SERVER_LOG"
+}
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+capture_cache_metrics
+trap capture_cache_metrics EXIT
+
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 438a3e9da..45dfde995 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7392,6 +7392,24 @@ qwen3.5-fp4-b200-sglang-agentic-mtp:
       - { tp: 2, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 14] }
       - { tp: 2, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 18, 20, 22, 24, 28, 32] }
 
+
+# Qwen3.8-Flash-Next NVFP4 AgentX on B200 via SGLang with native NEXTN MTP.
+# Day-zero recipe: SGLang is the plan-of-record engine for this model. TP1 per
+# the cookbook's verified single-node command; 126 GiB of NVFP4 weights fit on
+# one B200, so the model is not sharded.
+qwen3.8next-fp4-b200-sglang-agentic-mtp:
+  image: lmsysorg/sglang:qwen38flashnext
+  model: RadixArk/Qwen3.8-Flash-Next-NVFP4
+  model-prefix: qwen3.8next
+  runner: cluster:b200-nscale
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.8
+      search-space:
+      - { tp: 1, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
 qwen3.5-fp4-gb200-dynamo-sglang-agentic-mtp:
   image: lmsysorg/sglang:nightly-dev-20260818-c0b6474b
   model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ab92e7921..f82d613bf 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6576,3 +6576,17 @@
     - "Recipes sourced from srt-slurm (recipes/trtllm/qwen3.5-fp4/inferencex/gb300/{mtp,stp})."
     - "Runner: launch_gb300-nv.sh bumped from NVIDIA/srt-slurm@v1.0.29 to v1.0.72 for the dynamo-trt+qwen3.5+fp4 path."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2730
+
+- config-keys:
+    - qwen3.8next-fp4-b200-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add the day-zero Qwen3.8-Flash-Next NVFP4 AgentX recipe on B200 with SGLang native NEXTN MTP at TP4 and concurrency 1/4/8/12/16."
+    - "Serve RadixArk/Qwen3.8-Flash-Next-NVFP4 with modelopt_fp4 quantization, the trtllm_mha attention backend, and the flashinfer_trtllm MoE runner, following the Qwen3.5 NVFP4 B200 sibling."
+    - "Pin throughput runs to an interim acceptance length of 3.24 measured at three speculative tokens; eval-only runs keep real target verification."
+    - "Correct the serve flags to the SGLang cookbook's verified single-node command for this model: TP1 rather than TP4, the flashinfer linear-attention prefill and decode backends, bfloat16 Mamba SSM, and checkpoint-derived quantization."
+    - "Set the acceptance length to the committed golden thinking_on value of 2.32 at three speculative tokens, replacing the interim 3.24."
+    - "Keep the bfloat16 Mamba SSM state the cookbook specifies: SGLang requires it on SM100 or newer whenever the flashinfer linear-attention decode backend is selected."
+    - "Route the qwen3.8next fp4 prefix in the B200 nscale launcher to its staged checkpoint under the models tree; the launcher previously rejected the prefix outright."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2751
diff --git a/runners/launch_b200-nscale-compat.sh b/runners/launch_b200-nscale-compat.sh
index 6621bc8a8..9dea33302 100644
--- a/runners/launch_b200-nscale-compat.sh
+++ b/runners/launch_b200-nscale-compat.sh
@@ -92,6 +92,25 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
 elif [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH="/scratch/models/Kimi-K3"
     export SRT_SLURM_MODEL_PREFIX="kimik3"
+elif [[ $MODEL_PREFIX == "qwen3.8next" && $PRECISION == "fp4" ]]; then
+    # Staged on the compute nodes like every other model here, so this is the
+    # ordinary branch again: no hf download and no writable target needed.
+    # Verified on im-b200-c004: 126 GB, index present, 206/206 shards, no
+    # .incomplete leftovers. The candidate search mirrors the dsv4 branch so a
+    # differently named staging directory needs no code change.
+    SELECTED_MODEL_PATH=""
+    if [[ -n "${MODEL_PATH:-}" && -d "${MODEL_PATH}" ]]; then
+        SELECTED_MODEL_PATH="$MODEL_PATH"
+    else
+        for candidate in /scratch/models/Qwen3.8-Flash-Next-NVFP4 /scratch/models/Qwen3.8-Flash-Next; do
+            if [[ -d "$candidate" ]]; then
+                SELECTED_MODEL_PATH="$candidate"
+                break
+            fi
+        done
+    fi
+    export MODEL_PATH="${SELECTED_MODEL_PATH:-/scratch/models/Qwen3.8-Flash-Next-NVFP4}"
+    export SRT_SLURM_MODEL_PREFIX="qwen3.8next-fp4"
 else
     echo "Unsupported model prefix/precision: $MODEL_PREFIX/$PRECISION"
     echo "Available models under /scratch/models:"