diff --git a/.env.example b/.env.example
index 4788a31..d9e8183 100644
--- a/.env.example
+++ b/.env.example
@@ -22,20 +22,20 @@ UNCENSORED_PROJECTOR_FILE=qwen3.8-27b-abliterated/mmproj-Qwen3.8-27B-ABLITERATED
 VISION_PROJECTOR_FILE=qwen/mmproj-BF16.gguf
 
 FAST_CONTEXT=76800
-FAST_BATCH_SIZE=64
-FAST_UBATCH_SIZE=32
+FAST_BATCH_SIZE=2048
+FAST_UBATCH_SIZE=64
 MEDIUM_CONTEXT=160000
 MEDIUM_BATCH_SIZE=2048
 MEDIUM_UBATCH_SIZE=512
 LARGE_CONTEXT=192000
 LARGE_BATCH_SIZE=2048
-LARGE_UBATCH_SIZE=128
+LARGE_UBATCH_SIZE=256
 ULTRA_CONTEXT=262144
 ULTRA_BATCH_SIZE=2048
 ULTRA_UBATCH_SIZE=128
 UNCENSORED_CONTEXT=80000
 UNCENSORED_BATCH_SIZE=2048
-UNCENSORED_UBATCH_SIZE=128
+UNCENSORED_UBATCH_SIZE=256
 FAST_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 MEDIUM_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 MEDIUM_TENSOR_SPLIT=85,15
diff --git a/compose.yaml b/compose.yaml
index d5e59a9..b291b88 100644
--- a/compose.yaml
+++ b/compose.yaml
@@ -105,9 +105,9 @@ services:
       - --threads-batch
       - "${LLAMA_THREADS_BATCH:-6}"
       - --batch-size
-      - "${FAST_BATCH_SIZE:-64}"
+      - "${FAST_BATCH_SIZE:-2048}"
       - --ubatch-size
-      - "${FAST_UBATCH_SIZE:-32}"
+      - "${FAST_UBATCH_SIZE:-64}"
       - --parallel
       - "${FAST_PARALLEL_SLOTS:-1}"
       - --kv-unified
@@ -269,7 +269,7 @@ services:
       - --batch-size
       - "${LARGE_BATCH_SIZE:-2048}"
       - --ubatch-size
-      - "${LARGE_UBATCH_SIZE:-128}"
+      - "${LARGE_UBATCH_SIZE:-256}"
       - --parallel
       - "${LARGE_PARALLEL_SLOTS:-1}"
       - --kv-unified
@@ -431,7 +431,7 @@ services:
       - --batch-size
       - "${UNCENSORED_BATCH_SIZE:-2048}"
       - --ubatch-size
-      - "${UNCENSORED_UBATCH_SIZE:-128}"
+      - "${UNCENSORED_UBATCH_SIZE:-256}"
       - --parallel
       - "${UNCENSORED_PARALLEL_SLOTS:-1}"
       - --kv-unified
diff --git a/config/install.env.example b/config/install.env.example
index 0cc1426..2f0adaf 100644
--- a/config/install.env.example
+++ b/config/install.env.example
@@ -80,15 +80,15 @@ DEFAULT_REASONING_EFFORT=off
 # draft-model artifact is neither downloaded nor passed to llama-server.
 
 FAST_CONTEXT=76800
-FAST_BATCH_SIZE=64
-FAST_UBATCH_SIZE=32
+FAST_BATCH_SIZE=2048
+FAST_UBATCH_SIZE=64
 MEDIUM_CONTEXT=160000
 MEDIUM_BATCH_SIZE=2048
 MEDIUM_UBATCH_SIZE=512
 MEDIUM_TENSOR_SPLIT=85,15
 LARGE_CONTEXT=192000
 LARGE_BATCH_SIZE=2048
-LARGE_UBATCH_SIZE=128
+LARGE_UBATCH_SIZE=256
 LARGE_TENSOR_SPLIT=86,14
 ULTRA_CONTEXT=262144
 ULTRA_BATCH_SIZE=2048
@@ -96,7 +96,7 @@ ULTRA_UBATCH_SIZE=128
 ULTRA_TENSOR_SPLIT=80,20
 UNCENSORED_CONTEXT=80000
 UNCENSORED_BATCH_SIZE=2048
-UNCENSORED_UBATCH_SIZE=128
+UNCENSORED_UBATCH_SIZE=256
 UNCENSORED_TENSOR_SPLIT=90,10
 UNCENSORED_MTP_MAX=2
 LLAMA_THREADS=6
diff --git a/docs/PREFILL_BATCH_TUNING_20260915.md b/docs/PREFILL_BATCH_TUNING_20260915.md
new file mode 100644
index 0000000..a4c3c52
--- /dev/null
+++ b/docs/PREFILL_BATCH_TUNING_20260915.md
@@ -0,0 +1,22 @@
+# Prefill-Batch-Tuning vom 15. September 2026
+
+Alle fünf Qwen-Profile wurden auf Athena mit demselben kalten Textprompt von
+rund 54.000 Tokens getestet. Prompt-Cache-Treffer wurden ausgeschlossen. Als
+Zielwert gilt der schnellste stabile Lauf, nicht die größte startbare Zahl.
+
+| Profil | vorher | getestet | produktiv | Prefill vorher | Prefill produktiv |
+|---|---:|---:|---:|---:|---:|
+| Fast | 64 / 32 | 128/64, 2048/64, 2048/96, 2048/128 | **2048 / 64** | 806 tok/s | 1.176 tok/s |
+| Medium | 2048 / 128 | 2048/512, 2048/1024 | **2048 / 512** | 1.232 tok/s | 1.597 tok/s |
+| Large | 2048 / 128 | 2048/256, 2048/512 | **2048 / 256** | 1.231 tok/s | 1.542 tok/s |
+| Ultra | 2048 / 128 | 2048/256, 2048/512 | **2048 / 128** | 1.407 tok/s | 1.407 tok/s |
+| Uncensored | 2048 / 128 | 2048/256, 2048/512 | **2048 / 256** | 1.166 tok/s | 1.475 tok/s |
+
+Die Werte sind `Batch / Micro-Batch`. Fast OOMte mit Micro-Batch 96 während
+des langen Prompts und mit 128 beim Start. Ultra OOMte bereits beim Start mit
+256 und 512. Medium 1024 sowie Large und Uncensored 512 liefen, waren aber
+langsamer als der jeweils kleinere produktive Wert und benötigten mehr Reserve.
+
+Die Messung prüft Prefill und VRAM unter einem langen Textprompt. Änderungen an
+Modell, Kontextgröße, Vision-Projektor, Tensor-Split oder llama.cpp-Build
+erfordern einen neuen Vergleichstest.
diff --git a/platform/llama/README.md b/platform/llama/README.md
index 4d4871d..811acd3 100644
--- a/platform/llama/README.md
+++ b/platform/llama/README.md
@@ -49,3 +49,6 @@ sie alternativ als dokumentierte Referenz für vollständige systemd-Overrides
 verwendet werden.
 
 Der Profilwechsel erfolgt ausschließlich über `platform/scripts/llama-profile`.
+
+Die produktiven Batch- und Micro-Batch-Werte samt Messungen stehen in
+[docs/PREFILL_BATCH_TUNING_20260915.md](../../docs/PREFILL_BATCH_TUNING_20260915.md).