Running NVIDIA Cosmos Locally on Mac: Gets Motion Right, Fails at Counting
MaziyarPanahi · x · 2026-08-08
A developer ran NVIDIA's Cosmos-Reason1-7B model locally on a Mac Studio (M3 Ultra) using MLX 8-bit quantization, combined with Meta SAM 3.1 for segmentation.
He designed an interesting eval: asking the model to predict the next 3 seconds of video, then rolling the actual footage to let the model grade itself. The results show the model accurately predicted the motion trajectory of jets, but hilariously failed at counting (counting 6 objects when there were only 4).
Related event: Testing NVIDIA Cosmos: Accurate Trajectories but Fails at Counting(4 posts)→
More from Fun
- Meme: Cybertruck lifted by rocket thrusters celebrates 'AI that does all your work' — coherence · 2026-10-16
- Tokyo's Gemini AI Photo Booth turns visitors into plush toys and dioramas — bradneuberg · 2026-10-16
- Tinyfool sparks debate over whether Tesla Roadster 2 counts as 'useless' innovation — tinyfool · 2026-10-16
- Beff Jezos: when a cracked autist starts detailing their plans, believe them — beffjezos · 2026-10-16
- Tinyfool: does Xiaomi's foldable released days before iPhone count as innovation? — tinyfool · 2026-10-16
- Gemini users hit usage caps on 3.1 Pro and 4 Argon, resets not until Oct 16 — R3d-Gr33n-Blu3 · 2026-10-16