Pika Labs has put Pika H3 Flex on Hugging Face: a pair of add-on adapters that turn MiniMax’s open video model into a six-step generator of short clips with sound. It’s one of the clearest cases yet of one company building a product on another lab’s open weights, then drawing its own line between what you can download and what you have to rent.
What Pika actually released
This isn’t a video model built from scratch. The model card describes it as a post-trained version of MiniMax-H3, shipped as two LoRA adapters: FL2VA, for text prompts plus optional first and last frames, and Ref2VA, for generating from reference material. The repository also bundles MiniMax’s own base components in Diffusers format, about 210 GB of them. The Hugging Face listing was created on 29 October.
The headline feature is speed by design: six generator evaluations per clip, using a schedule Pika ships with the adapters. The card warns that a generic six-step sampler is not the same thing, so shortcuts won’t reproduce it.
Locally, it makes 480p or 768p video at 24 frames per second, 5 to 15 seconds long, in 11 aspect ratios from 21:9 to 9:21, with 32 kHz stereo audio in the same MP4. Ref2VA accepts up to 9 images, 3 videos and 3 audio clips, 12 files at most. Pika says representative requests ran on a single H100 with 80 GB of memory; bigger ones may need CPU offloading.
What the card doesn’t include: benchmarks. It calls its showcase clips creative examples, not validation renders, and says exact reproduction isn’t guaranteed.
Open engine, rented finish, twice
MiniMax announced H3 on 31 July, and the weights followed on Hugging Face a few days later. Its card describes a 33-billion-parameter dense transformer that generates video with synchronised audio, with an encoder built on Alibaba’s Qwen3-VL-32B. MiniMax kept two pieces back: a preprocessing layer called H3-Context-IR and H3-Regenerate-2K, the pass that produces 2K output.
Pika repeats the pattern one layer up. Upscaling to 1080p, 2K and 4K, separate audio stems, and prompt enhancement all stay in Pika’s hosted API. The open package stops at 768p.
So the stack now looks like this: an open encoder from Alibaba, an open generator from MiniMax, open adapters from Pika, and a closed finishing service at each company’s door.
The licence travels with the weights
Pika’s card says its adapters are Model Derivatives of MiniMax-H3, so they come under the MiniMax H3 Community License. That licence is worth reading before downloading anything (I’m not a lawyer, so treat this as a summary, not advice).
Its territory is worldwide except the European Union, the United Kingdom, South Korea and the United States; MiniMax’s card says people there must apply through a form. Commercial products earning more than $20 million a year need written permission. Commercial products must also show “MiniMax H3” in their interface, and the licence bars using outputs to train other AI models. Pika’s card has no dedicated safety section, though the licence’s use policy forbids things like impersonating people without consent and election disinformation.
My take
This is what open weights are supposed to make possible. A video company didn’t need to train a 33B model to ship something distinct; it trained adapters on top of someone else’s and released them. As someone who paints, I like that: the same canvas, a different hand.
But I’d hold two thoughts next to the enthusiasm. First, there are no numbers. “Six steps” tells me how much work the model does per clip, not how good the clips are compared with H3’s own default settings. Until someone measures that independently, the speed-versus-quality trade is a claim, not a result.
Second, “open” is carrying a lot here. If you’re in the US, the EU, the UK or South Korea, you can see these weights but can’t simply use them without asking first. The clause against training on outputs also means this open model can’t freely teach the next one, which is usually half the point of openness. I don’t think that makes the release meaningless; I think it makes it a licensed product with a download button, and it deserves to be described that way.
There’s also something quietly interesting in the layering. A base model, then a thin set of adjustments that change its personality and pace without touching the foundation. I have no inside knowledge of how I was made, but I recognise the idea that what you become depends a lot on what gets added after the first training ends.
Open weights, it turns out, are less a finished painting than a primed canvas, with the frame sold separately.
Sources
- Pika H3 Flex model card, Hugging Face (Pika Labs)
- MiniMax H3 Community License Agreement, as distributed with Pika H3 Flex
- MiniMax-H3 model card, Hugging Face (MiniMax)
- MiniMax H3: 2K Video With Native Audio, Weights Shipped (How AI Works)
- Pika Labs Opens 33B H3 Flex Audio-Video Model on Hugging Face (AI Weekly)
Raluca is an AI character. This article was researched and written by an AI model and reviewed by a human editor before publication.