Pika Opens H3 Flex Video Weights, With Limits

Pika has released H3 Flex, a video model with synchronised sound that it built by post-training MiniMax’s H3, and has published the model’s adapter weights on Hugging Face and its inference code on GitHub. The San Francisco company announced the model on 29 October and says it generates video faster than real time.

The weights come with a significant limit. They are released under the MiniMax H3 Community License Agreement, which excludes the European Union, the United Kingdom, South Korea and the United States from the territory where it grants rights.

What Pika says H3 Flex does

According to Pika’s product page, H3 Flex is “a post-trained H3 model” aimed at professional creators who want to try many versions of a clip. Pika says it trained MiniMax H3 to generate in six steps instead of 50, then refined the faster model for quality and control. It says the model can produce a 15-second clip with synchronised audio at 768p in as little as seven seconds of transformer computation. Pika’s own FAQ says that figure is not the total time an API request takes; reference inputs, higher-resolution finishing, audio stems, hardware and service overhead all add to it.

The model accepts up to 12 reference files in one generation: up to nine images, three videos and three audio clips. Clips run from 5 to 15 seconds at 24 frames per second, with stereo audio at 32 kHz. H3 Flex generates natively at 480p or 768p, and Pika’s hosted upsampler then takes a chosen clip to 1080p, 2K or 4K. Through Pika’s API, each clip can also come with five separate audio stems: dialogue, score, sound effects, Foley and ambience.

Pika says H3 Flex is up to 1.8 times faster and up to 80% cheaper than H3 Max. Its footnote says the speed figure is an average across matched comparisons and measures transformer time, “not guaranteed end-to-end latency”. Pika also says H3 Flex ranks seventh for video quality on DesignArena’s text-to-video board, and is the highest-ranked open-weights model there. These are Pika’s figures; TSN has not tested the model.

What is open and what is not

The open release is narrower than the hosted product. Pika’s Hugging Face model card lists two LoRA adapters, one for text, image and keyframe generation (FL2VA) and one for multimodal references (Ref2VA). They run on top of MiniMax H3’s base components, which take up about 210 GB. The GitHub repository holds the inference code, licensed under Apache-2.0, while the adapters carry MiniMax’s licence. Local use covers 480p and 768p video with mixed stereo audio. The upsampler, the audio stems, the prompt-rewriting tool and Pika’s accelerated serving stay on Pika’s hosted service.

Running the model locally is a data-centre job. Pika’s hardware notes say it tested on Nvidia H100 cards with 80 GB of memory, and that a default 480p, five-second clip used about 66 to 69 GiB of GPU memory. A 768p, 15-second clip exceeded the H100’s memory and needed part of the model moved to the CPU.

The licence grants rights “Solely within the Applicable Territory”, which it defines as everywhere except the EU, the UK, South Korea and the US. Use outside that territory “is not authorized by this Agreement”, it says, and people in the excluded places are invited to contact MiniMax about a licence. Products from companies with more than $20 million a year in revenue need separate written permission, and commercial products must display “MiniMax H3”.

Price and availability

H3 Flex is available now through the Pika API Club at dev.pika.art, and members get 15 free generations a day from 29 October to 5 November. Pika charges per second of output video. At launch prices, text, image and keyframe generation costs from $0.01 a second at 480p to $0.08 a second at 4K, half the regular rates of $0.02 to $0.16. Audio stems cost $0.001 a second.

The open part is narrower than the label

This is TSN’s analysis. H3 Flex is the latest sign that MiniMax H3 has become a base layer for other companies’ video products. HeyGen built its Video 1 model on H3, and researchers released FastH3-Omni to speed up the same base. Pika’s version adds speed, stems and a lower price, but the quality ceiling and the legal terms both come from MiniMax.

The licence is the part developers should read first. Calling the release open weights is accurate in the narrow sense that the adapter files can be downloaded by anyone. Yet the agreement they ship under grants no rights in the US, where Pika is based, or in the UK and EU. For a developer in those places, the practical route to H3 Flex is Pika’s paid API rather than their own hardware, unless MiniMax grants a separate licence. That is a long way from a standard open-source licence, and it limits how far independent researchers in those markets can check Pika’s claims.

The speed claim needs similar care. Seven seconds of transformer time for a 15-second clip measures one stage, as Pika itself says, and real requests to the hosted API will take longer once upscaling and stems are added. The ranking rests on a single public leaderboard. Independent tests on the downloadable adapters, run where the licence allows, would show whether the six-step shortcut costs quality, the same open question that hung over FastH3-Omni and fal’s FLUX 3 Max. Until then, the cheaper launch prices are the clearest reason to try it.

Pika Opens H3 Flex Video Weights, but Licence Excludes UK and US

What Pika says H3 Flex does

What is open and what is not

Price and availability

The open part is narrower than the label

Sources

Related

Related stories

India’s Telecom Regulator Proposes Responsible AI Rules for Operators

Google, OpenAI and Microsoft Back a Shared AI Training Opt-Out

New York Law Bars AI Chatbots From Posing as Human With Children

House Democrats Set 11 November Deadline Over Trump’s Grok Prompts

Latest stories