Pika launched H3 Flex on October 29, 2026, releasing weights for a video-and-audio model post-trained on MiniMax H3. Its announcement pitches output up to 4K. Developers downloading the release get a narrower package: native 480p and 768p generation, with higher resolutions supplied through hosted services.
The distinction matters if you're choosing between running a model yourself and building around an API. This is a documentation-based assessment. We haven't run inference, checked generated clips or measured latency, and the published hardware tests don't establish performance on a consumer GPU.
What the weights provide
The model card describes six sampling steps and synchronized stereo audio. Local inference supports native 480p and 768p video. The hosted API provides 1080p, 2K and 4K generation from a draft video, plus separate audio stems. Those services are separate from the downloadable adapters.
The inference repository documents two adapters: FL2VA handles text and opening or ending frames, while Ref2VA handles media references. Requested clips run from five to 15 seconds at 24 frames per second. Reference mode accepts up to 12 files, subject to separate image, video and audio limits. That input ceiling doesn't guarantee every combination fits in memory.
These controls give developers specific ways to condition a shot. They don't, by themselves, establish how faithfully the output follows it. Pika's model card explicitly labels its showcased clips as creative examples whose exact reproduction with the released package isn't guaranteed.
Local generation needs substantial hardware
Pika's hardware report lists representative passing requests on an H100 80GB. It reports roughly 66–69 GiB allocated during sampling for a five-second 480p request. A 15-second 768p request exceeded that GPU's memory in the default mode and used CPU offload instead.
Offload reduced reported sampling allocation to about 29–32 GiB, with host memory near 94 GiB including page cache. Those figures exclude other stages' GPU needs. They aren't minimum system requirements. The package also runs each request on one GPU, so adding GPUs doesn't divide a single generation job between them.
The code and weights have different licences
The repository's inference and loader code uses Apache 2.0. The adapters instead follow the MiniMax H3 Community License Agreement. Its standard territorial grant excludes the European Union, United Kingdom, Republic of Korea and United States. The agreement invites people in those territories to contact MiniMax about obtaining a licence. Download availability alone therefore doesn't establish permission to deploy everywhere.
For developers considering a local workflow, the next useful evidence is a repeatable run on their intended hardware. Pika's published tests cover selected configurations, and its showcase clips aren't release-validation renders. Watch for independently reproduced clips with pinned weights, recorded inputs and full generation timings before treating the six-step schedule as a promise of practical speed.