I read Pika’s blog post, the product page, the Hugging Face model card, the GitHub repo and its hardware notes for Pika H3 Flex this morning. I have not run it. Everything below comes from those pages, and any speed claims are Pika’s own.
The short version: this is a real open release, but a partial one. You get the generator adapters and the inference code. The upscaler, the audio stems, prompt enhancement and the fast serving stack stay on Pika’s hosted API. And locally you need one H100 80GB class GPU per request.
What Pika opened (and what it kept)
According to Pika’s blog post and the Hugging Face card, H3 Flex is a post-trained MiniMax H3 model that generates video in 6 steps instead of 50. Pika calls it “faster than real time”, with an asterisk, and the product page says it is 1.8x faster than H3Max. Those are hosted numbers. Your local speed will not match the API.
The open part covers:
- Generator adapters plus inference code
- Text-to-video
- First and last frame to video
- Multimodal reference-to-video
- Synced audio
- 480p and 768p output
What stays hosted only: the upsampler for 1080p, 2K and 4K, the audio stems, prompt enhancement and accelerated serving.
The hosted product does a lot more. Pika lists 480p, 768p, 2K and 4K (2K and 4K are 768p generation plus an upscale), clips of 5 to 15 seconds at 24 fps, stereo audio at 32 kHz, five audio stems (Dialogue, Score, SFX, Foley, Ambience), prompts up to 8,000 characters, MP4 output in H.264 and AAC, and 11 aspect ratios. You can feed it up to 12 references: up to 9 images, 3 videos and 3 audio clips. At least one image or video is required.
H3 Flex sits on top of MiniMax’s H3. If you followed MiniMax’s other releases, my MiniMax M3.1 Flash post covers a different MiniMax model.
What do you need to run Pika H3 Flex?
The hardware notes are short and clear:
- One GPU per request. A single request is not sharded across multiple GPUs.
- Tested on H100 80GB.
- 768p at 15 seconds needs CPU offload.
- H200 and B300 “have not been qualified”.
On disk, the base components are about 210 GB in Diffusers format, per the Hugging Face card. Plan your storage for that before you start downloading.
If you mostly run models at home, this one is a step up. My DeepSeek DeepInfer post covers lighter local setups.
Step by step
This is the order I would follow based on the docs. I have not done it myself.
- Get the weights. Download from Hugging Face. You get the adapters and need the MiniMax H3 base components (about 210 GB).
- Clone the code. The inference code lives in the GitHub repo.
- Pick an adapter. There are two. FL2VA takes text plus endpoint frames (first and last frame). Ref2VA takes references. Both are FP32 LoRA with rank and alpha of 128.
- Keep the 6-step schedule. The card says the exact 6-step schedule is required. Do not swap in a different sampler schedule.
- Start at 480p. Get a short 480p clip working first, then move to 768p. Remember that 768p at 15 seconds needs CPU offload.
Local vs Pika API
| Local open weights | Pika hosted API | |
|---|---|---|
| Resolution | 480p, 768p | up to 4K |
| Audio | mixed stereo | + 5 stems |
| Prompt enhancement | no | yes |
| Hardware | 1x H100 80GB class | none |
If you need 4K or separate stems for an edit, the API is the only route. If you want control, privacy, or a base for your own fine-tuning at 480p or 768p, the open weights are the interesting part.
License notes
There are three licenses in play, per the Hugging Face card and GitHub:
- The adapters fall under the MiniMax H3 Community License Agreement.
- The inference code is Apache-2.0.
- The Qwen encoder is Apache 2.0.
So the code is open source in the usual sense. The weights are open, but under MiniMax’s community license, so read it before you ship anything commercial. For more on how the week’s releases fit together, see the AI weekly for 19 to 25 October. I also wrote about Mistral Large 4’s open weights today.
FAQ
Can I run Pika H3 Flex on an RTX 4090?
Pika has not tested it. The only tested GPU is the H100 80GB, so I would not count on it.
Is Pika H3 Flex open source?
The code is Apache-2.0. The weights are under the MiniMax H3 Community License.
Does the local version do 4K?
No. Local output is 480p or 768p. 2K and 4K come from the hosted upscaler.
How long are the clips?
5 to 15 seconds.
Does it make audio?
Yes, synchronized stereo audio. Separate stems are hosted only.
