Add benchmark graphs to card

README.md

@@ -38,6 +38,18 @@ Which to pick: **q5_k** or **q5_0** for the best size and accuracy (byte-identic
 
 Only the large `ggml_mul_mat`-fed weights (Qwen3 and Whisper attention/FFN projections, the adaptor linears, and the token embedding, 343 tensors) are quantized; norms, biases, the conv stem, positional embeddings, and the mel filterbank stay F32.
 
+## Benchmarks
+
+Same audio, same F32 weights, same threads, byte-identical transcript. moss-transcribe.cpp (ggml, CPU) stays under real time where PyTorch does not, and the gap holds as clips get longer:
+
+![RTF vs audio length: moss-transcribe.cpp vs PyTorch on CPU](benchmarks/rtf_vs_length.png)
+
+Quantization makes the model both smaller and faster (the decode is memory-bandwidth bound), with the transcript byte-identical through q5:
+
+![Quantization ladder: size and speed by dtype](benchmarks/quant_ladder.png)
+
+Full methodology and the reproducible harness are in the [benchmarks](https://github.com/mudler/moss-transcribe.cpp/blob/master/benchmarks/BENCHMARK.md).
+
 ## Usage
 
 ```sh