Verification: add reproducible test image, inference snippet, and results

README.md

@@ -45,18 +45,41 @@ print(output)
 
 ## Verification (smoke test)
 
-This conversion was smoke-tested on Apple silicon with `mlx-vlm` after quantization.
-Both a text-only prompt and a vision prompt (on a deterministic generated image of
-coloured shapes + a text label) were run against the 8-bit weights.
+Smoke-tested on Apple silicon with `mlx-vlm` after quantization — a text prompt plus a
+vision prompt against a deterministic, network-free test image (coloured shapes + a text
+label), so the test is fully reproducible.
 
-**Text prompt** — _"In one sentence, what is the capital of France?"_
+### Test image
+
+![smoke test image](assets/smoke_test.jpg)
+
+### How it was run
+
+```python
+from mlx_vlm import load, generate
+from mlx_vlm.prompt_utils import apply_chat_template
+from mlx_vlm.utils import load_config
+
+model, processor = load("incept5/Qwen3.8-27B-MLX-8bit")
+config = load_config("incept5/Qwen3.8-27B-MLX-8bit")
+
+# vision
+msgs = [{"role": "user", "content": "Describe the shapes and their colours in this image, and read any text you see."}]
+prompt = apply_chat_template(processor, config, msgs, num_images=1)
+print(generate(model, processor, prompt, image=["test.jpg"], max_tokens=256, verbose=False))
+```
+
+(Text prompt used the same pattern with `num_images=0` and no `image=` argument.)
+
+### Results
+
+**Text** — _"In one sentence, what is the capital of France?"_
 
 ```
 The capital of France is Paris.
 ```
 
-**Vision prompt** — _"Describe the shapes and their colours in this image, and read any text you see."_
-(test image: red square top-left, blue circle top-right, green inverted triangle bottom-centre, label "MLX VISION TEST")
+**Vision** — _"Describe the shapes and their colours in this image, and read any text you see."_
 
 ```
 Based on the image provided, here is a description of the shapes, their colors, and the text present: