diff --git a/server.py b/server.py
index 19fba22..890010f 100644
--- a/server.py
+++ b/server.py
@@ -1820,8 +1820,8 @@ _STT_BACKEND_METRICS: dict[str, dict] = {
     "faster_whisper": {"speed": "~70× RT · GPU",    "latency": "0.5–2 s",     "quality": "large-v3",       "ram": "1.5 GB VRAM"},
     "whisper_cpp":    {"speed": "~8–15× RT · CPU",  "latency": "1–5 s",       "quality": "large-v3 Q5",    "ram": "~1 GB RAM"},
     "groq_whisper":   {"speed": "fastest cloud",    "latency": "0.5–1 s",     "quality": "Whisper Turbo",  "ram": "cloud · 0"},
-    "nvidia_parakeet":{"speed": "~200× RT · GPU",   "latency": "<0.3 s",      "quality": "Parakeet-TDT",   "ram": "~1 GB"},
-    "nvidia_router":  {"speed": "GPU routed",       "latency": "~0.5 s",      "quality": "varies",         "ram": "varies"},
+    "nvidia_parakeet":{"speed": "~200× RT · GPU",   "latency": "<0.3 s",      "quality": "Parakeet-TDT",   "ram": "~3 GB"},
+    "nvidia_router":  {"speed": "GPU routed",       "latency": "~0.5 s",      "quality": "varies",         "ram": "~11 GB"},
 }
 
 
@@ -3213,7 +3213,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Weak per-request style; instruct may be ignored.",
             "best_for": "Known voices, multilingual cloning, benchmarks, and reliable speaker identity.",
             "uses_wav": True, "style_aware": False, "true_streaming": False,
-            "speed": "~0.3× GPU", "latency": "1–3 s", "quality": "Premium clone", "ram": "~5 GB",
+            "speed": "~0.3× GPU", "latency": "1–3 s", "quality": "Premium clone", "ram": "~6 GB",
         },
         "voice_design": {
             "purpose": "Create or reuse prompt-designed voices from natural-language descriptions.",
@@ -3245,7 +3245,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Language and speaker are controlled by the backend voice config; per-request style text is usually ignored.",
             "best_for": "Fast local NVIDIA TTS voices and OpenAI-compatible assistant playback.",
             "uses_wav": False, "style_aware": False, "true_streaming": False,
-            "speed": "~0.05× GPU", "latency": "0.3–0.8 s", "quality": "High", "ram": "~1 GB",
+            "speed": "~0.05× GPU", "latency": "0.3–0.8 s", "quality": "High", "ram": "~10 GB",
         },
         "nvidia_zeroshot": {
             "purpose": "Clone a saved library voice through NVIDIA Magpie TTS Zeroshot NIM.",
@@ -3253,7 +3253,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Best with a clear 3-10 second prompt. Optional quality params can be configured in Settings.",
             "best_for": "Fast NVIDIA reference-audio cloning, streaming-class use cases, live agents, and games.",
             "uses_wav": True, "style_aware": False, "true_streaming": False,
-            "speed": "~0.1× GPU", "latency": "0.5–1 s", "quality": "High clone", "ram": "~1 GB",
+            "speed": "~0.1× GPU", "latency": "0.5–1 s", "quality": "High clone", "ram": "~10 GB",
         },
         "nvidia_flow": {
             "purpose": "Clone a saved library voice through NVIDIA Magpie TTS Flow NIM.",
@@ -3261,7 +3261,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Offline high-fidelity clone path; prompt transcript must match the reference audio.",
             "best_for": "Studio-style dubbing, narration, and podcast-quality offline generation.",
             "uses_wav": True, "style_aware": False, "true_streaming": False,
-            "speed": "~0.2× GPU", "latency": "1–2 s", "quality": "Studio", "ram": "~1 GB",
+            "speed": "~0.2× GPU", "latency": "1–2 s", "quality": "Studio", "ram": "~10 GB",
         },
         "kokoro": {
             "purpose": "High-quality English TTS with Kokoro 82M model. OpenAI-compatible endpoint.",
@@ -3277,7 +3277,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Text only — no voice ID or style parameters.",
             "best_for": "Lightweight local TTS on port 8027. Minimal setup, fast response.",
             "uses_wav": False, "style_aware": False, "true_streaming": False,
-            "speed": "fast", "latency": "0.2–1 s", "quality": "High", "ram": "varies",
+            "speed": "fast", "latency": "0.2–1 s", "quality": "High", "ram": "~4 GB",
         },
         "xtts": {
             "purpose": "XTTS v2 via xtts-api-server. OpenAI-compatible endpoint with speaker selection.",
@@ -3285,7 +3285,7 @@ def _backend_capabilities(backend: str) -> dict:
             "style": "Speaker selected by voice ID. Style instruction not supported.",
             "best_for": "Local multi-speaker TTS with XTTS v2 model. Coqui/daswer123 docker setup.",
             "uses_wav": False, "style_aware": False, "true_streaming": False,
-            "speed": "~0.3× GPU", "latency": "1–3 s", "quality": "High", "ram": "3–4 GB",
+            "speed": "~0.3× GPU", "latency": "1–3 s", "quality": "High", "ram": "~5 GB",
         },
     }
     return caps.get(_clean_preview_backend(backend), {})
diff --git a/static/sections/s-llms.html b/static/sections/s-llms.html
index 3fb991a..bf4ed6b 100644
--- a/static/sections/s-llms.html
+++ b/static/sections/s-llms.html
@@ -557,7 +557,7 @@ docker run -p 8880:8880 --gpus all \
               <span class="llm-metric-chip"><span class="mdi mdi-lightning-bolt"></span> ~0.5× GPU RTF</span>
               <span class="llm-metric-chip"><span class="mdi mdi-clock-outline"></span> 1–3 s</span>
               <span class="llm-metric-chip"><span class="mdi mdi-star-circle-outline"></span> High (WAV clone)</span>
-              <span class="llm-metric-chip"><span class="mdi mdi-memory"></span> 3–4 GB VRAM</span>
+              <span class="llm-metric-chip"><span class="mdi mdi-memory"></span> ~5 GB</span>
             </div>
             <p class="llm-local-desc">Coqui XTTS &mdash; multilingual voice cloning from a 6-second sample. 17 languages. Compatible with this app&rsquo;s voice library.</p>
             <div class="llm-local-url">