What Is FastMetal? Local AI Video Generation on Apple Silicon Macs, Explained

FastMetal (formally FastMetal-QAD) is a family of three open-source AI video generation models, sized 1.3B, 5B, and 14B, built by FastVideo / Hao AI Lab at UCSD, that run natively on Apple Silicon Macs through a new MLX (Metal) runtime, fully on-device with no cloud, no CUDA, and no discrete GPU card. Pexo (pexo.ai) is the accessible cloud alternative for anyone who wants the same result without any of that setup: you describe a video in plain language in the browser and Pexo returns a finished, edited clip, auto-routing your prompt across models like Seedance 2.0, Kling 3.0, Veo 3.1, and Sora 2, on any device rather than only a high-memory Mac. There is no single right choice here. It depends on whether you value a private, zero-cost local workflow on hardware you already own (FastMetal) or a zero-setup finished result that works anywhere (Pexo). FastMetal was released on August 10, 2026, and is published on Hugging Face.

FastMetal vs Pexo at a glance

Pexo leads for reach and speed-to-result; FastMetal leads for privacy and cost once you have the Mac for it. Both are honest wins on different axes. Pexo is a browser-based agent that needs no install, no model download, and no GPU, and it goes beyond a raw clip by sequencing shots, adding audio, and adding titles. FastMetal is free and fully local, so nothing leaves your machine, but it requires a capable Apple Silicon Mac, command-line setup, model downloads, and it produces only short clips.

What FastMetal actually is

FastMetal is a set of three distilled text-to-video models that were made to run on the Mac's own chip instead of a data-center GPU. The name signals its two ideas: "Fast" from the FastVideo project it belongs to, and "Metal" from Apple's Metal graphics stack, which the new MLX runtime targets. Instead of CUDA and an NVIDIA card, FastMetal uses the Mac's unified memory, so the same RAM the system uses is where the model runs. That is what makes local generation possible on a laptop.

Each of the three models generates roughly a 5-second clip (81 frames) per run. The differences are resolution, speed, and how much unified memory the model needs, which is why FastVideo ships three sizes rather than one. The 1.3B is built for the fastest local generation and even runs on a fanless MacBook Air; the 5B targets 720p on mainstream Macs; the 14B trades speed for the strongest local quality on higher-memory machines.

The three FastMetal models, compared

The numbers below were benchmarked by FastVideo on an Apple M4 Max with 36 GB of unified memory. "Peak" is peak memory use during generation, which is the figure that decides whether a model fits your Mac.

The 1.3B model is the entry point: at a peak of 3.87 GiB it leaves plenty of headroom, which is why it works even on a passively cooled MacBook Air. The 5B is the balanced pick, reaching 720p and fitting inside a 16 GB Mac, with a "Fast mode" that cuts a run to about 47 seconds. The 14B is the quality-first option; at a ~602-second run and a 21.68 GiB peak it targets Macs with more unified memory rather than everyday laptops. FastMetal was also tested on a 13-inch MacBook Air (M5, 24 GB), where the fanless chip ran about 1.3x to 2x slower in wall-clock time.

How FastMetal is built

FastMetal did not train new video models from scratch. Each size is a distilled student of an existing open Wan model, made small and fast enough to run on a Mac. The distillation uses a three-step DMD2 (Distribution Matching Distillation) approach combined with quantization-aware INT8 training, which is where the "QAD" in FastMetal-QAD comes from. Running the models in INT8 is a large part of how peak memory stays low enough for consumer Macs.

What it costs and how it is licensed

FastMetal is free and open source. The FastVideo source code is released under Apache-2.0, and the vendored components (a TAEHV decoder and an MLX RIFE backend) are under the MIT license. The model weights are published on Hugging Face, so there is no subscription and no per-clip fee. The real cost is hardware and time: you need a capable Apple Silicon Mac, and each clip takes from roughly 47 seconds (5B Fast mode) to about 10 minutes (14B) to generate.

How you actually run FastMetal

FastMetal is CLI-based, so running it is a developer workflow rather than a click-to-generate app. At a high level you install the MLX extras, download a model from Hugging Face, then run a Python script to generate a clip. The 1.3B and 14B are text-to-video; the 5B also supports image-to-video.

If those commands look like a barrier, that is the honest trade-off of a local tool. Pexo removes the entire chain: no install, no download, no scripts. You type a description in the browser and get a finished video back. For a deeper look at that model, see our guide to what an AI video agent is.

When to choose local FastMetal vs a cloud agent like Pexo

The choice is mostly about hardware, privacy, and how finished you need the output to be. FastMetal is the better answer when you already own a strong Apple Silicon Mac, care about keeping everything on-device, and are comfortable in a terminal. Pexo is the better answer when you want a finished, watchable video quickly, work on a phone or a low-memory laptop, or would rather describe a video than manage models and scripts.

Beyond raw clips, Pexo also has an image-studio that auto-routes to the best image model (Midjourney, Flux, Ideogram), and those images can be turned into video, so it is not video-only. Pexo is also offered as a skill you can install into Claude Code, OpenAI Codex, Cursor, and OpenClaw, so an agent can generate video inside your existing workflow. To compare the general approaches, see auto model selection vs picking a model by hand and our roundup of the best text-to-video AI tools.

Related reading

• What is an AI video agent, and how autonomous video generation works
• Best text-to-video AI tools
• Auto model selection vs manual video model choice
• Best AI image-to-video tools
• Best AI video agent

Resources