diff --git a/backend/services/fit.py b/backend/services/fit.py
index 8ddd44a..30fce7d 100644
--- a/backend/services/fit.py
+++ b/backend/services/fit.py
@@ -81,7 +81,9 @@ def max_ctx_in_budget(params_b: float, quant: str, budget_gb: float) -> int:
     ctx_budget = budget_gb - weights
     if ctx_budget <= 0:
         return 2048
-    per_8k = (max(params_b, 7) / 7) * 0.8
+    # KV pro 8k — EXAKTE Inverse von estimate_memory_gb (sqrt, kalibriert an
+    # Hermes-14B@128K≈19GB). Vorher linear → für große Modelle viel zu konservativ.
+    per_8k = (max(params_b, 7) / 7) ** 0.5 * 0.84
     raw_ctx = (ctx_budget / per_8k) * 8192
     best = _NICE_CTX[0]
     for c in _NICE_CTX: