diff --git a/src/whichllm/cli.py b/src/whichllm/cli.py
index 9db55ce..ff81f01 100644
--- a/src/whichllm/cli.py
+++ b/src/whichllm/cli.py
@@ -127,7 +127,9 @@ def _auto_min_params_for_profile(hardware: HardwareInfo, profile: str) -> float
         return None
     if not hardware.gpus:
         return 2.0  # CPU-only: tiny is the only practical choice
-    usable_ram = int(hardware.ram_bytes * 0.80)
+    from whichllm.hardware.memory import estimate_usable_ram
+
+    usable_ram = estimate_usable_ram(hardware.ram_bytes)
     best_vram_gb = max(
         (usable_ram if g.shared_memory and g.vram_bytes == 0 else g.vram_bytes)
         for g in hardware.gpus
diff --git a/src/whichllm/engine/compatibility.py b/src/whichllm/engine/compatibility.py
index e0676b5..632965f 100644
--- a/src/whichllm/engine/compatibility.py
+++ b/src/whichllm/engine/compatibility.py
@@ -8,6 +8,7 @@
 from whichllm.engine.quantization import estimate_weight_bytes
 from whichllm.engine.types import CompatibilityResult
 from whichllm.engine.vram import estimate_vram
+from whichllm.hardware.memory import estimate_usable_ram
 from whichllm.hardware.types import GPUInfo, HardwareInfo
 from whichllm.models.types import GGUFVariant, ModelInfo
 
@@ -56,8 +57,7 @@ def check_compatibility(
 
     vram_required = estimate_vram(model, variant, context_length)
 
-    # Reserve 20% of RAM for OS and other processes
-    usable_ram = int(hardware.ram_bytes * 0.80)
+    usable_ram = estimate_usable_ram(hardware.ram_bytes)
 
     # Determine best GPU
     best_gpu: GPUInfo | None = None
diff --git a/src/whichllm/hardware/memory.py b/src/whichllm/hardware/memory.py
index dd4ffde..082277c 100644
--- a/src/whichllm/hardware/memory.py
+++ b/src/whichllm/hardware/memory.py
@@ -13,6 +13,17 @@ def detect_ram_bytes() -> int:
     return psutil.virtual_memory().total
 
 
+def estimate_usable_ram(total: int) -> int:
+    """Estimate RAM available for model loading after OS/background reserve.
+
+    Uses a bounded-reserve formula: total - clamp(total * 0.15, 4 GiB, 32 GiB).
+    """
+    _GiB = 1024**3
+    reserve = int(total * 0.15)
+    reserve = max(4 * _GiB, min(reserve, 32 * _GiB))
+    return max(0, total - reserve)
+
+
 def detect_disk_free_bytes(path: str | None = None) -> int:
     """Get free disk space in bytes at the given path.
 
diff --git a/tests/test_compatibility.py b/tests/test_compatibility.py
index 0610775..f170f21 100644
--- a/tests/test_compatibility.py
+++ b/tests/test_compatibility.py
@@ -1,6 +1,7 @@
 """Tests for compatibility checking."""
 
 from whichllm.engine.compatibility import check_compatibility
+from whichllm.hardware.memory import estimate_usable_ram
 from whichllm.hardware.types import GPUInfo, HardwareInfo
 from whichllm.models.types import GGUFVariant, ModelInfo
 
@@ -92,7 +93,7 @@ def test_shared_memory_amd_apu_uses_system_memory_pool():
 
     assert result.can_run is True
     assert result.fit_type == "full_gpu"
-    assert result.vram_available_bytes == int(hw.ram_bytes * 0.80)
+    assert result.vram_available_bytes == estimate_usable_ram(hw.ram_bytes)
     assert not any("offload" in w.lower() for w in result.warnings)
     assert not any("cpu only" in w.lower() for w in result.warnings)
 
@@ -121,7 +122,7 @@ def test_windows_shared_memory_amd_apu_does_not_emit_rocm_warning():
 
     assert result.can_run is True
     assert result.fit_type == "full_gpu"
-    assert result.vram_available_bytes == int(hw.ram_bytes * 0.80)
+    assert result.vram_available_bytes == estimate_usable_ram(hw.ram_bytes)
     assert not any("rocm" in w.lower() for w in result.warnings)
     assert not any("offload" in w.lower() for w in result.warnings)
 
diff --git a/tests/test_memory.py b/tests/test_memory.py
new file mode 100644
index 0000000..879acf9
--- /dev/null
+++ b/tests/test_memory.py
@@ -0,0 +1,40 @@
+"""Tests for hardware.memory — estimate_usable_ram bounded-reserve formula."""
+
+import pytest
+
+from whichllm.hardware.memory import estimate_usable_ram
+
+_GiB = 1024**3
+
+
+def _expected_usable(total: int) -> int:
+    reserve = int(total * 0.15)
+    reserve = max(4 * _GiB, min(reserve, 32 * _GiB))
+    return total - reserve
+
+
+@pytest.mark.parametrize(
+    "total_gb",
+    [16, 32, 64, 128, 1024],
+    ids=["16GB", "32GB", "64GB", "128GB", "1TB"],
+)
+def test_estimate_usable_ram(total_gb):
+    total = total_gb * _GiB
+    assert estimate_usable_ram(total) == _expected_usable(total)
+
+
+def test_16gb_hits_min_reserve():
+    total = 16 * _GiB
+    assert estimate_usable_ram(total) == total - 4 * _GiB
+
+
+def test_1tb_hits_max_reserve():
+    total = 1024 * _GiB
+    assert estimate_usable_ram(total) == total - 32 * _GiB
+
+
+def test_midrange_uses_percentage():
+    total = 64 * _GiB
+    expected_reserve = int(total * 0.15)
+    assert 4 * _GiB < expected_reserve < 32 * _GiB
+    assert estimate_usable_ram(total) == total - expected_reserve