Phosphene 5.0: Free Mac app generates AI video, images, voices and songs locally

What's going on

On November 8, a community developer announced Phosphene 5.0 on Reddit. The free app is released under the MIT license, a very permissive open-source license that also allows commercial use. Phosphene runs AI models for video, images, speech and music locally on Macs with Apple chips (M1 through M4). Everything is generated on your own machine, not on a cloud provider's servers. Under the hood, the app is built on MLX, Apple's machine learning library. MLX is designed around the unified memory of M-series chips: the CPU and GPU share a single memory pool, which is why larger models can run.

Earlier versions had to be installed through Pinokio, a one-click installer for AI apps. Version 5.0 brings its own Mac installer.

What's new in version 5.0

The key changes, according to the developer:

• Mac installer: Download it, open it, and setup runs automatically. No Terminal required. If you already use Phosphene via Pinokio, you don't have to download the models again. Pinokio still works.
• Redesigned Studio: You pick the model first and then see only the settings that apply to it.
• Video with sound: With LTX-2.5 (from Israeli company Lightricks) and Hailuo H3, you can create clips from text (text-to-video) or from a starting image (image-to-video). Setting up H3 is now automatic in the Mac app.
• Images: According to the developer, the P-H3-Image and FLUX 3 Klein models render text in images exactly as you type it, something image generators struggled with for a long time. Since version 5.0.3, the app generates images about five times faster by default on his M4 Max.
• Speech and lip sync: A line of text becomes a spoken voice, and the app then animates a face so the lips match the audio.
• Songs with vocals: In the "Music Studio," the YuE2 model turns your own lyrics into finished songs with vocals.
• Training your own characters: With a LoRA, a small add-on module you train on just a few images, you can teach LTX a specific character. That character then looks consistent across multiple clips.
• Lower memory requirements: Video and songs work with as little as 16 GB of RAM, character training from 24 GB, and Hailuo H3 from 36 GB.
• MCP server and CLI: Through MCP (Model Context Protocol), an open standard that lets AI assistants operate external tools, programs like Claude Code or Codex can generate clips on their own.

How fast is it? The developer's numbers

The developer benchmarked on his MacBook with an M4 Max and 64 GB, using the same prompt and the same seed each time. He ran each setting three times and reports the median:

These are the developer's own figures on one of the most powerful Mac chips. On an M1 or M2 with 16 GB it takes considerably longer, though the post doesn't say how much. The developer says more benchmarks are available on the project site, phosphene.video.

Why local video generation matters

Tech outlet Startup Fortune had already covered an earlier Phosphene version, still based on LTX 2.3, and sees it as part of a bigger trend. Cloud AI video services charge per second of video, by resolution or by compute time. If you experiment a lot and generate dozens of variations of a clip, the costs add up fast. Locally, each attempt costs only electricity and time. According to Startup Fortune, that also changes how you work: you experiment more because there's no price tag on every attempt.

Then there's privacy. Prompts and outputs never leave your machine, which can be a real selling point for commercial projects or product prototypes. Startup Fortune argues this opens up product ideas for small dev teams that would be too expensive with paid APIs. The outlet draws a line from Stable Diffusion for images, through llama.cpp and Ollama for language models, to video. Video is following the same pattern, just a year or two later, because the models are bigger and need more memory. Lightricks' LTX models are tuned for efficiency rather than top benchmark scores, which makes them well suited to local use.

What you need to try it

• Hardware: A Mac with an Apple Silicon chip (M1, M2, M3 or M4). Intel Macs, Windows and Linux aren't supported, since the app is built on MLX.
• RAM: At least 16 GB for video with LTX-2.5 and for songs. 24 GB if you want to train your own characters. 36 GB or more for Hailuo H3. More memory also helps at higher resolutions.
• Storage: Models are downloaded during setup. The post doesn't give exact sizes, but plan on several dozen gigabytes of free space to be safe.
• Installation: Download the Mac installer from the project site or GitHub and run it. Pinokio remains an alternative.
• Patience: Even on an M4 Max, a 5-second clip takes one to nine minutes. On smaller Macs, expect much longer waits.

Our take

Phosphene isn't a new AI model. It's a front end that bundles existing open models, and that's exactly what makes it interesting for Mac owners. Until now, local video generation was really only convenient on Nvidia GPUs. With the installer and support starting at 16 GB, the barrier to entry drops significantly. Having video, images, voice, lip sync and music in one app is handy if you want to produce short clips entirely offline. Tinkerers will also like the MCP integration: an AI assistant can use it to create entire series of clips automatically.

That said, all performance figures come from the developer, and there are no independent tests yet. How the results stack up against cloud services like Veo, Sora or Kling remains to be seen. Wait times for video generation on Macs are also still long. Startup Fortune described an earlier version of the project as "early and still rough around the edges." Hands-on testing will show whether version 5.0 has changed that much. One more thing to keep in mind: the MIT license covers the app itself, but the bundled models come with their own licenses, which may restrict commercial use.