Pika released H3 Flex, an open-weight video model post-trained on MiniMax H3 that generates a 15-second clip with soundtrack in as little as 7 seconds, according to Pika's launch post. Pika built it for professional iteration, with drafts at 768p, a finishing mode that upscales to 4K, and API pricing from $0.01 per second.
- Faster and cheaper than H3 Max, per Pika. Pika says H3 Flex is up to 1.8x faster than H3 Max and up to 80% cheaper. The speed figure is averaged across matched comparisons and measures DiT time, not end-to-end latency.
- Drafts at 768p, finishes at up to 4K. You explore ideas at lower resolution, then upscale to 1080p, 2K or 4K in a finishing mode.
- Up to five exportable audio stems. Dialogue, score, SFX, Foley and ambience can export as separate tracks.
- Open weights from launch. Pika published the code and weights on day one, with the model listed as a 33-billion-parameter image-to-video model built with Diffusers.
- Top open-weight model on DesignArena, per Pika. Pika says H3 Flex is the highest-ranked open-weight model on DesignArena's text-to-video board.
Up to 12 references, prompt edits and multi-shot clips in a single generation
H3 Flex accepts up to 12 reference inputs, including sketches, voice samples, images, sounds, video and moodboards, according to Pika's product page. Editing starts from an existing video, where a reference or a prompt can add, remove or swap elements and adjust characters, camera moves, looks and voices.
A single generation can hold a 15-second sequence with multiple shots, hard cuts, dialogue, music and sound design. Pika lists complex camera moves, motion graphics, product films and animated shorts among the genres it targets, with orbits, dollies, cranes, whip pans and lens changes directed in natural language.
Pika says the audio stems let you isolate and refine any facet of a clip's soundscape.
$0.01 per second on Pika's API, with code and weights open from launch
Pika, which runs an AI creative platform for video, images and audio, offers H3 Flex through its Pika API Club with prices starting at $0.01 per second. Pika is giving away 15 generations per day during a two-week launch window.
Inference code is in Pika's GitHub repo. Pika says it made the model open source from day one, in keeping with a design aimed at giving creators flexibility and speed.
The weights sit on Pika's Hugging Face page, where the model is listed as a 33-billion-parameter image-to-video model built with Diffusers.
Pika says H3 Flex ranks seventh in video quality on DesignArena's text-to-video board with a score of 1293, above base H3, H3 Max and Seedance 2.5.
H3 Flex joins fal's H3 Max, VDN-H3 and other post-trained MiniMax H3 models
H3 Flex is built on H3 from MiniMax, an open, general-purpose omni-modal generation model that understands text, images, video and audio.
MiniMax's H3 launch post says the base model generates video with native stereo audio at up to 2K and 15 seconds; H3 Flex adds a finishing mode that upscales output to 4K. Other models post-trained or adapted from H3 include the following.
H3 Max by fal. fal Research post-trained it from the open H3 weights, and fal says it generates a 5-second video in under 3 seconds, roughly 35x the throughput of the official H3 endpoint.
H3 Max Turbo by fal. This is the configuration fal tuned hardest for throughput. It bills at half of H3 Max at every resolution, starting at $0.025 per second at 480p, and fal's own measurements show a 5-second 480p clip in 0.44 seconds.
VDN-H3 by OpenVDN. This hybrid-attention version adds a linear-attention branch and two LoRA adapters to H3. It generates a 14.4-second clip in about 9 seconds end-to-end on 8 B200 GPUs, and FreeVideo, built on VDN-H3, runs H3 on GPUs with as little as 8 GB of VRAM.
Other speed options. MiniMax's H3 integrations index also lists Turbo, PDD and FastH3 as speed options for H3.
Hugging Face lists more than 200 fine-tuned models built on MiniMax H3.
Rankings vary by board. Pika claims H3 Flex outranks H3 Max on DesignArena's text-to-video board, while fal says H3 Max ranks first on DesignArena's image-to-video board.
