Refresh card with final measurements

README.md

@@ -35,6 +35,7 @@ list is libc, libm and libgomp.
 | Checkpoint → package | 2.23 GB → **774 MB** (3.03x) |
 | Bits per weight | 5.28 |
 | Tensors compressed | 16 |
+| Resident memory | **870 MB** on CPU, 919 MB on GPU — 1.12x the package |
 | Architecture | 16 blocks, GQA 32q/8kv, gated short convolutions |
 
 ## Quality
@@ -46,6 +47,21 @@ list is libc, libm and libgomp.
 | **Result** | **no detectable cost** (95% CI [0.9899x, 1.0104x], t = +0.02) |
 
 
+## Running it
+
+Measured on a Jetson AGX Thor: CPU is 12 threads of a 14-core Arm part, GPU is
+the integrated Blackwell. Decode is the median of five runs; prefill is from a
+162-token prompt. Both backends produce **byte-identical** output.
+
+| | CPU (12 threads) | GPU |
+|---|---|---|
+| decode | 2.43 tok/s | 24.66 |
+| prefill | 16.7 tok/s | 90.0 |
+
+Load takes 1.34 s on CPU and 0.82 s on GPU,
+the latter including the host-to-device upload.
+
+
 65,408 paired tokens of FineWeb-Edu in **127 independent
 512-token windows**. Both models score identical tokens and are
 compared per token, which cuts the standard error 15.7x versus