On October 29 Pika released Pika H3 Flex, and what matters if you use Pika H3 Flex in ComfyUI is easy to say: video with synchronized audio in 6 steps instead of 50, with open weights and three ready-made workflows in the template library. Anyone generating video with AI knows the problem well: every test costs minutes (or credits), and the right shot shows up after ten attempts. A model that shortens every attempt changes the way you work, not just the timing.
Here's our take. We've been working with generative AI since 2024 and today we use ComfyUI on a perfectly ordinary 12 GB RTX 3060, so when we read "open source" the first question is always: can I really use it, here, on my card and for my projects? With Pika H3 Flex the honest answer has two parts. The tech is promising. The license, if you work from Italy, is the real obstacle, and none of the articles we've read says so clearly. We haven't tried it yet: here you'll find the official data, checked against the sources.
Quick facts
| What it is | A "distilled" version of MiniMax H3 (retrained to give the same kind of result with far fewer steps): videos up to 15 seconds with stereo audio, in 6 steps |
| Who made it | Pika (Mellis, Inc.), starting from the MiniMax H3 model |
| Video memory required | Not stated. The most used reduced base model weighs about 21 GB (INT8) or 16 GB (w6a8) |
| Does it run on a 12 GB RTX 3060? | Not stated. On paper the files don't fit in 12 GB: part of the model would have to be moved to system RAM. Not measured by us |
| License | MiniMax H3 Community License for the weights; Apache 2.0 only for the example code |
| Commercial use from Italy | Not with the public license: the European Union is excluded. You need a separate license from MiniMax |
| Where to download/try it | Hugging Face Pika-Labs/Pika-H3-Flex and official ComfyUI templates; in the Pika app from November 2 |
| Free or paid | Free weights; upscaling up to 4K, separate audio tracks and prompt enhancement only paid on the Pika service |
What is Pika H3 Flex and how is it different from MiniMax H3?
Pika H3 Flex is MiniMax H3 retrained by Pika to generate in 6 steps instead of the original model's 50, keeping video and audio in the same generation. According to the official announcement, it creates clips from 5 to 15 seconds at 480p or 768p, 24 frames per second, with synchronized stereo audio.
In practice it's not a brand-new model: it's two LoRA (small files that change the behavior of an existing model) to apply on top of MiniMax H3, which we already covered in our article on MiniMax H3 Singularity V2 in ComfyUI. The two LoRA are:
- FL2VA: video from text, from a first frame, from a last frame or from both;
- Ref2VA: video from references, up to 12 files (9 images, 3 videos, 3 audio clips), useful to keep a character consistent, transfer a motion or swap a subject.
The technical paper explains that the step reduction was achieved with a distillation technique called PDMD, then refined with two more "taste" LoRA (ReFL and DPO) already merged into the result. On DesignArena, a leaderboard based on user votes, Pika H3 Flex entered at 7th place in text-to-video and is the top open-weight model. The comparisons published on the blog, on the other hand, were made by Pika itself: take them as manufacturer data. Pika also admits the limits: physics, consistency across multiple scenes and control of complex references still need work.
Does it really generate "faster than real time"?
Yes, but only in lab conditions and only for part of the computation. The line going around social media, "15 seconds of video in 7 seconds", refers, in a footnote of the announcement, to the transformer computation alone (the model's "brain") for a 768p clip, measured on servers with eight B300 GPUs. It's not the total time, and it's not the time you'll see on your PC.
The figure closest to a "creator" card comes from the official ComfyUI tutorial: a 5-second clip at 1344×768 in about a minute "on a single card", without saying which one. A tester quoted by ComfyUI Wiki got a 2560×1440 video in 3 minutes and 17 seconds, with a first 6-step pass, an upscale and a second 4-step pass, but without stating the hardware.
Also watch out for what stays on Pika's servers: upscaling to 1080p, 2K and 4K, splitting the audio into five tracks (dialogue, music, effects, foley, ambience) and automatic prompt enhancement are not in the open weights. Locally you get 480p and 768p with a single stereo audio track.
How do you use Pika H3 Flex in ComfyUI?
You use it with the three official templates in the Video section of the ComfyUI template library: video_pika_h3_flex_t2v (text only), video_pika_h3_flex_fl2va (first and last frame) and video_pika_h3_flex_ref2va (references). The text template requires ComfyUI 0.43.2, released on October 30; the other two work from 0.30.0. If you're just starting out, our beginner's guide to ComfyUI explains how to update and load a template.
The files are the same as the existing MiniMax H3 templates, plus the two new LoRA (about 2 GB each in BF16): the reduced base model in INT8 (a compressed format that roughly halves the size compared to BF16) goes in models/diffusion_models, the LoRA in models/loras, the Qwen3-VL text encoder in models/text_encoders and the two VAE (the decoders that turn the result into images and sound) in models/vae.
Don't touch the settings at random: LCM sampler with manual sigmas, CFG 1.0, LoRA strength 1.0. The tutorial warns that adding another acceleration LoRA (Turbo or Lightning) "breaks" the result, and that a regular sampler set to 6 steps doesn't reproduce Pika's schedule. Another honest note: in the README Pika says it verified the file conversion, but not a full render in ComfyUI. Some differences from their engine are possible.
How much VRAM does Pika H3 Flex need?
Nobody has said. Neither Pika nor ComfyUI states the VRAM (the graphics card's memory) required; the model card on Hugging Face mentions tests on an 80 GB H100, and the base components in Diffusers format take up about 210 GB on disk.
We can only do the math on the files: the reduced base model weighs about 21 GB in INT8 and about 16 GB in the w6a8 version, plus the text encoder. On a 12 GB card like ours, then, ComfyUI would have to move part of the model to system RAM ("offload"), with longer times and the need for a lot of memory in your PC. There are projects like FreeVideo that claim to run MiniMax H3 with 8 GB, but nobody has shown a combination with the Flex LoRA yet. If you're considering an upgrade, our guide to the best GPUs for AI walks through the memory question.
Can you use Pika H3 Flex for work in Italy?
Not with the public license, not even for local personal use. The weights are distributed under the MiniMax H3 Community License, the same as the base model: in the text included in the repository, the "excluded territories" are the European Union, the United Kingdom, South Korea and the United States, and using the weights or their outputs outside the permitted territory "is not authorized".
We're sorry to write this, because on paper it's one of the most interesting open video models right now. But "open weights" doesn't mean "free for everyone". The license says that anyone in the excluded territories can ask MiniMax for a separate license; outside the excluded territories, commercial use is allowed below 20 million dollars in annual revenue, displaying the "MiniMax H3" notice. Pika's online service has its own terms, which need to be read separately. If the topic interests you, our article on AI video locally or online covers the pros and cons of both routes.
Try it right away
Before downloading dozens of gigabytes, we recommend these steps:
- Read the license: open the LICENSE file in Pika's repository and check whether your use case is covered. From Italy, without an agreement with MiniMax, it isn't.
- Watch the launch video "Introducing Pika H3 Flex" on Pika's YouTube channel to get an idea of the real quality before investing time.
- Check your disk space: between base model, text encoder, VAE and LoRA you need several dozen GB free on an SSD.
- If you can use it, update ComfyUI to 0.43.2, open the video_pika_h3_flex_t2v template and start with a 5-second clip at 480p: leave sigmas, CFG and LoRA strength unchanged, and only change the prompt.
To compare it with alternatives without territorial restrictions, our handbook of AI video models in ComfyUI collects licenses and requirements of the main models.
Other news
- Faster MiniMax H3 decoding: on October 28 the latent upscaler for H3 and support for Sol-H3-VAE landed in the ComfyUI code; according to the model card, it decodes 5 seconds of 768p video in 5 seconds instead of 12.5 on an RTX 5090 (GitHub). The changes were merged after version 0.43.1.
- StableGen 0.4.0: the free add-on (GPL-3.0) that textures 3D models in Blender using ComfyUI, released on October 28, adds editing of a painted area, automatic rigging and Qwen Image 2.1 (GitHub).
Frequently asked questions
Is Pika H3 Flex open source?
It's more accurate to say "open weights". Pika has released the model files and the inference code, but the weights follow the MiniMax H3 Community License, which excludes some territories and sets commercial conditions. Only the example code is Apache 2.0. Upscaling, separate audio tracks and prompt enhancement stay on the paid service.
What's the difference between Pika H3 Flex and MiniMax H3?
Pika H3 Flex uses MiniMax H3 as its base and adds two LoRA that cut the steps from 50 to 6, with extra training on video and audio quality. In ComfyUI the base model files are the same: what changes are the LoRA and the sampler settings, which should be left as in the official templates.
Does Pika H3 Flex reach 4K locally?
No. Locally it generates at 480p and 768p; upscaling to 1080p, 2K and 4K is a Pika service. In ComfyUI you can upscale with other tools, for example the latent upscaler for H3 just added to the code, but the result isn't the same as Pika's engine.
When is it coming to the Pika app?
According to the October 29 announcement, Flex comes to Pika's product from November 2, as well as via API. Prices aren't given in the announcement and should be checked on the official page. The online service has its own terms of use, different from the weights license.
If you want to keep up with the models you can actually use, you'll find all our articles in the ComfyUI section and hands-on videos on the YouTube channel @tutorial3d, with free workflows to download.
