Pika H3 Flex is here to try in Phosphene, the free, open-source Mac app by Mr Bizarro for making AI video with sound on your own Mac. This is a preview, and we're sharing it early because we'd love people to try it and tell us what breaks. We'll fix problems fast.
Pika Labs published its fast, six-step version of H3 on October 29, 2026, for NVIDIA H100 GPUs only. Phosphene runs it natively on Apple Silicon with our own MLX code. As far as we know, this is the first Mac port.
Pick "H3 Flex" in the engine picker at the top, right next to Hailuo H3. It works with Text-to-video and Image-to-video using a first frame, with dialogue and sound included like normal H3.
It's about 2.2x faster than Best at the same size. Measured on an M4 Max with 64 GB, for each 5-second clip:
- Draft (640x384): about 5 minutes.
- High (1024x576): about 16 minutes. Best took 36.5 minutes on the same shot and seed.
Here's a 15-second ping-pong rally in an empty pool, made with Image-to-video at High. It's three 5-second pieces joined together.
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/02_h3flex_rally_15s_high.mp4
The same shot and seed, side by side at High: Flex on the LEFT (6 steps, 16 min), Best on the RIGHT (15 steps, about 36 min).
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/04_flex_vs_best_same_seed.mp4
A woman at a kitchen table says a line. This Text-to-video clip took about 5 minutes at Draft.
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/03_h3flex_kitchen_draft.mp4
To try the preview:
- In Pinokio, press Update, then restart Phosphene.
- You need Hailuo H3 already installed.
- Pick "H3 Flex" in the engine picker at the top. The same click downloads its 2.8 GB adapter from Pika's own Hugging Face page and prepares it on your Mac. The picker shows the progress.
Known limits:
- A 15-second clip is three 5-second pieces stitched together. Later pieces can drift: a dog's breed or a sofa's colour might change.
- Full size (1344x768) runs; full timing to come (one test pass suggested about 37 minutes per 5-second clip).
- Pika's second mode, for transferring a reference or motion into a video, is not included yet.
Tip: name every person who should stay in the scene in your prompt. The model drops people the prompt doesn't mention. Pika's own sitcom showcase does this too.
One small change to Best at High: its existing speed-up now uses the setting it was trained for, following an update to H3.
Please try the preview and leave feedback in the comments here or in GitHub issues. Tell us what works and what breaks so we can keep improving it.
Powered by MiniMax H3. Pika H3 Flex by Pika Labs (Mellis, Inc.). The weights are a MiniMax H3 Model Derivative under the MiniMax H3 Community License; Pika's code is Apache-2.0. We download the weights from Pika's Hugging Face at a fixed version, never re-host them, and convert the file on your Mac with a modification notice marking it as modified. The H3 licence excludes the EU, UK, South Korea and US, just as it already does for H3 in Phosphene.
Mac weights on Hugging Face: https://huggingface.co/Mrbizarro/Pika-H3-Flex-MLX
https://github.com/mrbizarro/Phosphene/releases/tag/v4.19.15
Pika H3 Flex is here to try in Phosphene, the free, open-source Mac app by Mr Bizarro for making AI video with sound on your own Mac. This is a preview, and we're sharing it early because we'd love people to try it and tell us what breaks. We'll fix problems fast.
Pika Labs published its fast, six-step version of H3 on October 29, 2026, for NVIDIA H100 GPUs only. Phosphene runs it natively on Apple Silicon with our own MLX code. As far as we know, this is the first Mac port.
Pick "H3 Flex" in the engine picker at the top, right next to Hailuo H3. It works with Text-to-video and Image-to-video using a first frame, with dialogue and sound included like normal H3.
It's about 2.2x faster than Best at the same size. Measured on an M4 Max with 64 GB, for each 5-second clip:
Here's a 15-second ping-pong rally in an empty pool, made with Image-to-video at High. It's three 5-second pieces joined together.
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/02_h3flex_rally_15s_high.mp4
The same shot and seed, side by side at High: Flex on the LEFT (6 steps, 16 min), Best on the RIGHT (15 steps, about 36 min).
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/04_flex_vs_best_same_seed.mp4
A woman at a kitchen table says a line. This Text-to-video clip took about 5 minutes at Draft.
Watch the clip with sound: https://github.com/mrbizarro/Phosphene/releases/download/v4.19.15/03_h3flex_kitchen_draft.mp4
To try the preview:
Known limits:
Tip: name every person who should stay in the scene in your prompt. The model drops people the prompt doesn't mention. Pika's own sitcom showcase does this too.
One small change to Best at High: its existing speed-up now uses the setting it was trained for, following an update to H3.
Please try the preview and leave feedback in the comments here or in GitHub issues. Tell us what works and what breaks so we can keep improving it.
Powered by MiniMax H3. Pika H3 Flex by Pika Labs (Mellis, Inc.). The weights are a MiniMax H3 Model Derivative under the MiniMax H3 Community License; Pika's code is Apache-2.0. We download the weights from Pika's Hugging Face at a fixed version, never re-host them, and convert the file on your Mac with a modification notice marking it as modified. The H3 licence excludes the EU, UK, South Korea and US, just as it already does for H3 in Phosphene.
Mac weights on Hugging Face: https://huggingface.co/Mrbizarro/Pika-H3-Flex-MLX
https://github.com/mrbizarro/Phosphene/releases/tag/v4.19.15