Upload README.md with huggingface_hub

README.md

@@ -36,17 +36,19 @@ underneath.
 
 ## What comes out
 
-<video controls muted playsinline width="512"
-  src="https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf/resolve/main/samples/corgi_512x288_4step.mp4">
-</video>
+![A corgi in a chef hat over a pan, four-step render](https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf/resolve/main/samples/corgi_512x288_4step.gif)
 
 *"A corgi in a chef hat flipping a pancake, sizzling sounds and a cheerful bark."*
 — 512×288, 39 frames at 24 fps, seed 42, **four steps**, nothing but the prompt.
-The audio is not a second model: the same transformer denoises both streams in
-one packed sequence. Also in the repository as
-[`samples/`](https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf/tree/main/samples):
-the AVI `cortiq animate` actually wrote, its `.wav`, and this mp4 (remuxed for
-the browser — the runtime itself never touches ffmpeg).
+
+The GIF is silent; the audio is the point, so take the
+**[mp4](https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf/resolve/main/samples/corgi_512x288_4step.mp4)**.
+It is not a second model: the same transformer denoises both streams in one
+packed sequence, on two different flow schedules.
+[`samples/`](https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf/tree/main/samples)
+also holds the AVI `cortiq animate` actually wrote and its `.wav` — the mp4 and
+the GIF are remuxes for the browser, and the runtime itself never touches
+ffmpeg.
 
 The LoRA is not a separate download: it is merged into the weights, so the file
 IS the 4-step model.