Local video generation is moving from a workstation hobby to a laptop feature, and the fastest proof shipped this week under the FastMetal name. FastMetal-QAD, an open-source family from the FastVideo team at UCSD’s Hao AI Lab, brings 1.3B, 5B, and 14B video models to Apple Silicon through a native MLX runtime, with the 5B model generating a clip in under a minute on a 16GB Mac.
What FastMetal Is
FastMetal-QAD is a series of open-source video generation models released this week, with three sizes aimed at different hardware: a 1.3B model for the smallest machines, a 5B sweet spot for mainstream Macs, and a 14B model for the high end. All three run natively on Apple Silicon through a newly written MLX runtime that moves the diffusion transformer and decoder onto the Metal GPU, so nothing needs to go through a cloud API.
The models are distilled using DMD2 into three-step student models, which is how a 5B model can produce video in under a minute instead of the multi-minute runs typical of larger diffusion models. They also use quantization-aware training on an affine INT8 grid, meaning the models are trained at the precision they ship with, avoiding the quality loss that normally comes from quantizing a model after the fact. Each model ships as both a Diffusers safetensors checkpoint and an MLX checkpoint.
The family name carries a lineage: the FastVideo team previously released FastWan-QAD in June, which brought video generation to a single consumer RTX 5090. FastMetal is the same approach pointed at Apple hardware, and the quantization-aware distillation technique is what makes both possible on consumer-class machines.
What the Numbers Say
The team published measurements from an Apple M4 Max with 36GB of unified memory, using three-step DMD generation, INT8 DiT, and the TAEHV decoder, with end-to-end time covering prompt encoding, model loading, denoising, and decoding export:
| Model | Fast mode | Notes |
|---|---|---|
| FastMetal-1.3B | fastest | runs on small machines |
| FastMetal-5B | ~47 seconds | the mainstream recommendation |
| FastMetal-14B | slower, higher quality | for high-end Macs |
The 5B model in fast mode produces a video in around 47 seconds, which matches the “under a minute” claim. One practical detail: the first generation pays a one-time cost because the umT5 text encoder takes about 47 seconds for a cold-start prompt encoding on the 5B model, but subsequent generations with the same prompt hit a content-addressed cache and jump straight to denoising. So the first clip is the slowest, and batch work with repeated prompts is much faster than the headline number suggests.
The team also re-tested on a 13-inch MacBook Air with an M5 chip, 24GB of RAM, 10-core GPU, and no fan. Both the 1.3B and 5B models run at the same resolution and in the same modes as on the Mac Studio, with output quality matching and peak memory differing by only a few hundred MiB. The cost is speed: generations take 1.3 to 2 times longer on the Air, which is an acceptable trade for a fanless laptop.
The Memory Math
Apple’s marketing RAM numbers are decimal, so a 24GB Mac actually has about 22.35 GiB, minus what the operating system takes. The 5B model peaks under 11 GiB in all modes, which means 720p generation is comfortable on a 16GB machine, not just a 24GB one. That is the number that matters for most buyers: video generation has gone from needing a workstation with an 80GB GPU to fitting in a base-tier MacBook Pro.
The model zoo is worth knowing before you download: the 1.3B and 5B models are from the Wan2.1 lineage, while the 14B model is larger and aimed at quality over speed. All of them work with the same MLX runtime, and installation is a standard Python package plus a Hugging Face download command.
The Honest Caveats
“Under a minute” is fast mode at 5B, and quality in fast mode is a trade against the slower modes the runtime also offers. The team lists several composable generation modes, and the 47-second number is the fast path, not the quality path. Users who want the best frames will pay more time per clip.
The measurements are the team’s own, from two Apple machines. Real-world results vary with thermals, background load, and memory pressure, and the MacBook Air’s fanless design means sustained batches will throttle. The 16GB claim applies to the 5B model specifically; the 14B model needs more headroom.
There is also the ecosystem maturity question. This is a young project, the MLX runtime is new, and the models are released days ago. Expect rough edges in prompt handling, tooling, and integration with existing video workflows, and do not treat it as a drop-in replacement for a polished commercial product yet.
Who Should Care
Mac users who have been watching local video generation from the sidelines get the real thing now: a 5B model that runs on a 16GB machine and produces a clip in under a minute. Creators who prototype video ideas will find the speed useful for iteration, even if final renders still go to a bigger machine. Developers building MLX-based video tooling get a new reference implementation. And anyone comparing local generation options should put FastMetal next to the other open-weights path on Apple Silicon: our MiniMax H3 on MLX review covers the 115GB multimodal model that set the previous bar for local video on a Mac, and the contrast between an 11GB fast model and a 115GB flagship is exactly the trade space worth understanding.
The Bottom Line
FastMetal-QAD is the strongest sign yet that local video generation is becoming a laptop feature rather than a workstation privilege. The 5B model’s under-a-minute generation on 16GB of RAM, backed by quantization-aware distillation and a purpose-built MLX runtime, is a genuinely useful capability today, not a research demo. The caveats are the usual young-project ones: team-measured numbers, fast-mode quality trade-offs, and rough edges. But the direction is unmistakable: if you have a recent Mac, you can now generate video locally, and it costs nothing per clip. That is a milestone worth paying attention to.