beam lane: 384K context (13.6x concurrency) + needle-in-haystack at 131K and 327K tokens (both correct, 60s/175s TTFT)

recipes/beam-1-0-nvfp4-4x/NOTE.md

@@ -121,6 +121,21 @@ Measured 2026-10-21 against the lane as served (DFlash2 k=7, FLASHINFER, fp8 KV)
   tool-call parsing is `--tool-call-parser beam` with
   `--enable-auto-tool-choice`; `--reasoning-parser beam` splits the channels.
 
+### Long-context needle test (the shape of a real long agent session)
+
+Served now with `--max-model-len 393216` → KV pool **5,337,742 tokens**,
+13.57x concurrency at the full window.
+
+| prompt | needle depth | TTFT | total | answer |
+|---|---|---|---|---|
+| 130,871 tok | 85% | 60.1 s | 61.2 s | ✅ `ARGON-7741` |
+| **327,054 tok** | 85% | 174.5 s | 176.0 s | ✅ `ARGON-7741` |
+
+327K is exactly the context size seen on the prod lane's longest real session —
+Beam retrieves a fact buried at 85% depth there without error, at ~1,875 tok/s
+prefill. Prefill latency (~0.5 s per 1K tokens) is the real cost of long
+sessions on this lane: keep prefix caching hot and the conversation incremental.
+
 ### Depth sweep (DFlash2 k=7, 2 runs/point, prefix caching on)
 
 | depth | pp t/s (ctx / pp2048) | tg t/s c1 | tg t/s c2 |

recipes/beam-1-0-nvfp4-4x/recipe.toml

@@ -54,9 +54,11 @@ args = [
     "--enable-auto-tool-choice",
     "--tool-call-parser", "beam",
     "--reasoning-parser", "beam",
-    # conservative first cut: 262144 of the model's 1M window. The KV pool is
-    # sized by the memory fraction below; raise once the pool is measured.
-    "--max-model-len", "262144",
+    # 384K of the model's 1M window: the measured KV pool is 5.27M tokens, so
+    # this still allows ~13 concurrent full-length requests, and it leaves room
+    # for the bench depth suite's 256K point (a 264K prompt does not fit a
+    # 262144 cap — 2026-10-21 depth run).
+    "--max-model-len", "393216",
     "--gpu-memory-utilization", "0.85",
     "--max-num-seqs", "8",
     "--trust-remote-code",