Why Lua Is Mac-Only: Apple Silicon Performance for AI Video Editing

Every time you edit a reel in Lua, the AI runs entirely on your Mac — not on a server somewhere else. Transcription, the raw cut, caption generation, b-roll placement, everything. This is unusual. Most AI video tools today process your footage in the cloud because AI models are expensive to run and most consumer hardware cannot keep up. But Lua was built from the ground up for one kind of machine: a Mac with Apple Silicon.

This decision — Mac-only, Apple Silicon only — is the most consequential one we have made. This post covers why.

What Apple Silicon gives AI that other hardware does not

Apple Silicon chips (M1, M2, M3, M4, M5 and beyond) share a design that is unusually good for local AI. The two features that matter most are unified memory and the Neural Processing Unit (NPU).

Unified memory means the CPU, GPU and NPU all access the same pool of RAM. Data does not have to be copied between separate video RAM and system RAM as it does on a PC with a discrete NVIDIA card. For a video-editing AI pipeline — read a frame, transcribe audio, generate a caption overlay, encode the result — this eliminates the biggest bottleneck that kills performance on traditional hardware.

The NPU is a dedicated on-chip processor designed for the matrix math that neural networks rely on. When Lua runs a Whisper transcription model or a caption-layout model, the NPU handles the heavy floating-point work while the GPU renders the timeline and interface. That parallelism is what makes on-device AI feel immediate rather than laggy.

On a MacBook Pro with 16 GB of RAM or more, these two features let Lua run a full AI editing pipeline at speeds that approach cloud latency — without sending a single byte over the internet.

The three things Lua does locally (and why each needs Apple Silicon)

Lua runs three AI-heavy workloads on your machine. Each one took advantage of Apple Silicon differently when we built it.

Transcription. Lua transcribes every take using an on-device Whisper model that has been optimised for Apple’s NPU via the llama.cpp Metal backend. A 15-second take transcribes in about two seconds on an M2 and under one second on an M4. The video never leaves the machine — only short audio chunks go to the model. This is the privacy promise that cloud transcription cannot match.

The raw cut. Once Lua knows every word and its timing, it compares overlapping takes of the same line, keeps the cleanest delivery, and removes retakes, pauses and false starts. This needs to happen fast because creators import 10 to 20 takes at once. Apple Silicon’s multi-core CPU handles take comparison across parallel threads, and the unified memory means there is no data-transfer delay between the comparison logic and the video decoder.

Caption styling and rendering. Making a single caption overlay involves: laying out text in English, Tamil, or Tanglish; positioning it to avoid the speaker’s face; computing word-level timing; and rendering the result into the video frame. Tamil script in particular requires complex glyph shaping that generic caption libraries get wrong. Lua runs its own caption engine that uses Apple Silicon’s GPU for rasterisation and the NPU for word-boundary detection. The result is frame-accurate captions in any of the three languages without a round trip to a cloud rendering service.

Why not a cloud app?

A cloud-based AI video editor would have been easier to build. Upload a video to a server, run the model on a GPU cluster, download the result. That is what most competitors do. We chose the harder path for three reasons.

Privacy. Your talking-head footage is often unpolished and personal — half-formed thoughts, imperfect takes, unreleased ideas. It should not live on a server. With Lua, the footage stays inside the Mac. The only data that leaves your machine is the export file you choose to share.

Latency. Cloud-based video editors require an upload step that takes minutes before any AI processing can start. Then the model runs, the result downloads, and the creator waits again. On-device processing starts the moment you click Next. The first time a creator experiences instant transcription on a Mac, the cloud approach feels like a relic.

Offline capability. Lua works without an internet connection. A creator in a co-working space with spotty Wi-Fi, on a flight, or in a rural area with slow broadband can still edit reels at full speed. This is a practical advantage that matters more than most benchmarks suggest.

What about Windows or Intel Macs?

We get this question often. The short answer is that Apple Silicon is not just faster than Intel Macs for our workload — it makes local AI feasible in a way that Intel Macs cannot match. An Intel Mac with no NPU and separate system and GPU memory would need to page data constantly, making on-device Whisper transcription and frame-accurate caption rendering impractically slow.

Windows PCs with high-end NVIDIA GPUs can, of course, run AI models very fast. But supporting Windows would mean maintaining a completely separate codebase for CUDA and dealing with a fragmented hardware landscape where no two machines offer the same AI capabilities. We chose to go deep on one platform rather than shallow on many.

That platform is the Mac with Apple Silicon, because it is the only consumer machine where every unit ships with a capable NPU, unified memory, and enough baseline RAM to run modern AI models locally.

What this means for creators

If you edit reels on a Mac with Apple Silicon, here is what the Mac-only decision means for your workflow:

• No subscription per video. Cloud tools charge per minute of processed footage. Lua runs on your hardware, so the cost is fixed.
• Your footage never leaves your machine. No upload bars, no waiting for server queues, no wondering where your raw takes ended up.
• Offline-first. Lua works without Wi-Fi. Every AI feature — transcription, raw cut, captions, b-roll — runs locally.
• Your speed scales with your hardware. An M4 MacBook Pro processes reels about 3x faster than an M1. As Apple Silicon improves, Lua gets faster with no update needed.

Try it on your Mac

Lua runs on any Mac with Apple Silicon and macOS 14 or newer. Download the app, drop in a few takes, and watch the AI process the entire reel on your machine. Visit trylua.com to get started.

If you want to understand more about how the raw cut works, read the deep dive on the Lua blog.