NVIDIA's 4B World Model Now Runs on a Mac. It Is Not a Camera.

On November 7, a developer posting as NetProjects put up an MLX port of NVIDIA's Cosmos3-Edge. The full build is 7.6 GB. The 8-bit build is 5.0 GB. On an M5 Max with 128 GB of unified memory, the 8-bit version ran NVIDIA's own image-to-video example, 832 by 480 at 121 frames, in 113 seconds, with a memory peak of 18.8 GB.
Read those numbers again in the units that matter. 113 seconds to generate about five seconds of video at 480p. That is roughly one frame per second of compute, on a laptop, for a model with four billion parameters.
The obvious reaction is that this is useless for making video. The correct reaction is that nobody built it for making video.
What Cosmos3-Edge actually is
Cosmos3-Edge is a 4B parameter world model. It generates 256p or 480p video, 50 to 150 frames, from text or from a starting image. It is released under the OpenMDW 1.1 license, which permits commercial use.
The important word is world model. NVIDIA built this family for robotics and autonomous driving, where the job is to predict what happens next given a scene and an action. The output is a hypothesis about physics: if the vehicle keeps this trajectory, does the pedestrian step out, does the crate slide, does the arm collide with the shelf.
NVIDIA's own model card says outputs "should not be treated as physically accurate simulation." That sentence is the whole design brief. The model is supposed to produce motion that is plausible enough to train against and wrong in ways you can measure. A photorealistic render that violates conservation of momentum is worse than an ugly one that respects it, if your downstream consumer is a policy network.
So the 480p ceiling is not a limitation to apologize for. Robotics training does not need 4K. It needs many variations, fast, cheap, and consistent enough that a policy trained on them transfers to the real world.
Why the Mac port matters
Because it changes who can run one.
Until now, evaluating a world model meant a GPU allocation. Small, but real, with a driver stack, a container, and somebody's cluster. That is a filter. It means the people who can poke at a model are the people who already have infrastructure, and the people who have the interesting robotics question often do not.
A 5.0 GB 8-bit build that runs on Apple silicon under MLX removes the filter. Anybody with a recent Mac and enough memory can generate a hundred clips overnight, look at them, and form an opinion. That is how a research direction becomes a field rather than a demo.
The 128 GB number is doing work here. 18.8 GB of memory peak means the machine does not have to be the top of the line. A 32 GB Mac should fit the 8-bit build with room for the OS. The 113 second figure will get worse on slower memory bandwidth, but the ability to run it at all is the threshold that matters.
There is also a subtler point about MLX specifically. Apple's framework is optimized for unified memory, which suits this workload. A 4B diffusion-style video model is memory bandwidth bound more than compute bound. That is why the port is plausible in the first place, and why it performs better than a naive estimate based on teraflops would suggest.
The caveats are real
This is an unofficial port. The author states plainly that it is not affiliated with NVIDIA. There is no demo. When the briefing site tried to load the GitHub install page, it failed.
A 5 GB download, an MLX environment, and a model with no demo and an install page that may or may not render is a weekend artifact from someone who wanted to see if it would work. It is not a product. Treat any performance claim accordingly, including the 113 seconds, which is a single data point on a single machine with a single example.
There is also a licensing detail worth reading before you build anything commercial on it. OpenMDW 1.1 permits commercial use, but it is not Apache 2.0 and not MIT. Read the terms, particularly if you are planning to distribute model weights or outputs.
And the usual quality caveat for world models. Feed it a photo of a kitchen and a prompt about a robot arm reaching for a cup, and you will get motion that reads as plausible. Look closely and the cup may change shape, the arm may pass through the counter, the lighting may drift. Those failures are exactly what a robotics researcher wants to measure. They are exactly what a video editor cannot use.
What 113 seconds buys you
Let me be concrete about the trade, because the number sounds bad and probably is not.
If you are generating training data for a manipulation policy, you do not need twenty clips. You need thousands, varied across lighting, object pose, camera angle and surface material. A pipeline that produces one 121-frame clip every two minutes on a laptop produces about 700 clips in a day, unattended. That is a usable dataset for a small experiment, and it costs electricity.
Compare that to renting GPU time for the same volume. The rental is faster per clip and requires a whole operational layer. For a graduate student or a two-person startup testing whether a world model is even the right approach, the laptop wins on time to first result, which is the only metric that matters at that stage.
The other thing 113 seconds buys you is honesty about scale. When each clip costs two minutes of your own machine's attention, you look at them. You notice the failures. When each clip costs four cents and arrives in a batch of five hundred, you aggregate and move on. Slowness is a forcing function for actually reading your outputs.
The pattern to watch
What matters this week is the timing. The port happened six days after the model's release cycle, done by an individual, for free, on hardware NVIDIA does not sell.
The diffusion path for open weights looks like this now. A lab releases weights under a permissive-ish license. Within a week, somebody has an MLX build, somebody else has a GGUF quantization, and a third person has a ComfyUI node. The official tooling arrives months later, if at all.
For NVIDIA this is mostly good. Cosmos is a strategic bet on world models becoming the training substrate for physical AI, and a substrate gets more valuable the more people can run it. Locking it to datacenter GPUs would protect a margin and shrink an ecosystem.
For Apple it is a quiet win. Every one of these ports is an argument that unified memory is enough for inference workloads that people assumed needed a discrete card. The M5 Max does not beat an H100 at anything. It does make a category of model runnable by people who will never touch an H100.
Whether the 113 seconds number survives contact with more users is an open question. Somebody will run it on an M4 Pro with 24 GB, hit a memory wall, and post a correction. That is fine. The useful claim is that a 4B world model now fits on a laptop, which changes who gets to have an opinion about world models. The speed is a footnote.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.