Update README.md

README.md

@@ -17,4 +17,16 @@ For more details, please refer to our ||coming soon release blog post||. Or try
 
 ## Usage
 
+Load the model and prepare it for inference. We use [FlexAttention for inference](https://pytorch.org/blog/flexattention-for-inference/), so calling `.compile()` is critical for fast decoding. Our `compile` implementation also handles warmup, so you can start making requests directly once it returns.
+
+```
+    moondream = AutoModelForCausalLM.from_pretrained(
+        "moondream/moondream3-preview",
+        trust_remote_code=True,
+        dtype=torch.bfloat16,
+        device_map={"": "cuda"},
+    )
+    moondream.compile()
+```
+
 * TODO: Add usage examples