News · SI
Pika releases H3 Flex, a fast video model with open weights
Pika has launched H3 Flex, a post-trained MiniMax H3 video model that generates clips with synchronized audio in six sampling steps and takes up to 12 references. Its generator adapters and inference code are open, while 4K upscaling and audio stems stay in Pika's hosted API.
Pika has launched H3 Flex, a video model it post-trained from MiniMax H3 to generate faster and take more direction[1]. It produces video with synchronized stereo audio, accepts up to 12 reference files, and ships with open weights so developers can run and modify it themselves[1]. Flex is available through the Pika API now and arrives in Pika's main creation product on 2 November[1].
01 / 04
Six steps instead of 50
Pika trained Flex to generate in six sampling steps instead of the 50 its base model uses, then refined the faster model for quality and control[1]. The company says it can make a 15-second 768p clip with audio in as little as seven seconds of transformer computation, though it notes that's a controlled benchmark rather than end-to-end API latency[1]. At launch, Flex ranked first among open-weights models and seventh overall on DesignArena's text-to-video leaderboard[1].
02 / 04
Built around references
Flex takes up to 12 references in one generation, mixing images, video and audio[1]. Pika's examples combine character sheets, sketches, product photos and motion-reference videos, and show reference-guided editing that swaps a setting or prop while keeping the original performance and timing[1]. Through the Pika API, users can also pull five separate audio stems (dialogue, score, sound effects, Foley and ambience) and upscale a chosen 480p or 768p draft to 1080p, 2K or 4K[1].
03 / 04
What's open, and what isn't
The open release is Flex's generator adapters plus inference code, covering text-to-video, first- and last-frame generation and reference-to-video at 480p and 768p[1]. The adapters are LoRA weights applied on top of MiniMax H3 base components, which take up about 210 GB, and Pika says representative requests ran on a single H100 80GB GPU[2]. The upsampler, stem separation, prompt enhancement and accelerated serving stack remain Pika-hosted features[1].
The license matters too. Flex is covered by the MiniMax H3 Community License, whose grant excludes the European Union, the United Kingdom, South Korea and the United States[3]. Businesses whose products earn more than US$20 million a year need separate written authorization from MiniMax[3].
04 / 04
Why it matters
Video models have mostly been closed services, and the open ones have been slow to iterate with. Flex brings a leaderboard-ranked model with native audio down to a handful of sampling steps and puts the weights on Hugging Face. The catch is the territorial license and the 4K and stem features held back for Pika's own API, so for many teams it's a model to experiment with locally and pay for at the finishing stage.
FAQ
Frequently asked questions.
01What is Pika H3 Flex?
Pika H3 Flex is a video generation model that Pika built by post-training MiniMax H3. It generates video with synchronized stereo audio in six sampling steps instead of 50, accepts up to 12 reference files, and was announced on 29 October 2026.
02How fast is it?
Pika says Flex can produce a 15-second 768p clip with synchronized audio in as little as seven seconds of transformer computation. That figure comes from a controlled benchmark and doesn't include total API latency, references, upscaling or service overhead.
03What exactly is open?
Pika released Flex's generator adapters and inference code on Hugging Face and GitHub. They cover text-to-video, first- and last-frame generation and reference-to-video with synchronized audio at 480p and 768p. The 1080p, 2K and 4K upsampler, audio-stem separation, prompt enhancement and Pika's accelerated serving stack stay hosted-only.
04Can I use the weights commercially?
The weights fall under the MiniMax H3 Community License. Its grant excludes the European Union, the United Kingdom, South Korea and the United States, and businesses whose products earn more than US$20 million a year need prior written authorization from MiniMax. Check the license before building on it.
05What hardware does it need to run locally?
The base components take up about 210 GB, and Pika says representative 480p and 768p requests ran on a single H100 80GB GPU. Larger requests may need CPU offload and a lot of system RAM.
That’s the story
Building with what’s new?
Spectrum Web Co ships products, places talent, builds social narratives and builds custom business tools as monthly services.
Related stories