news

Pika open-sources H3 Flex as two LoRA adapters, and the open files stop at 768p

Pika says the post-trained MiniMax H3 model sits seventh overall and first among open weights on DesignArena's text-to-video board at 1293 Elo, but the 4K finish, the stem separation and the prompt rewriter stay on Pika's servers.

By Daniel Ochoa · 31 October 2026

Pika released H3 Flex on October 29, 2026 and called it open source from day one, but the files on Hugging Face and GitHub cover 480p and 768p generation. The 4K output in the launch headline comes from a hosted upscaler, the five separate audio stems come from hosted stem separation, and Pika's prompt rewriter is not in the download either. The downloadable part and the benchmarked part are not the same product.

The model itself is not a new architecture. H3 Flex is a post-train of MiniMax H3, distributed as two task adapters that load on the 33B base, under the MiniMax H3 Community License rather than a permissive one, with the example inference code under Apache 2.0. Pika's own framing is that it is MiniMax H3 tuned on curated video, audio and reference-based tasks covering camera movement, character consistency, motion transfer and editing, then re-optimised around a six-step sampling process instead of the base model's much longer schedule.

1293 Elo, seventh overall, 16 points above the model it was built from

Pika's placement claim is specific and checkable: seventh for video quality on DesignArena's text-to-video board with a score of 1293, first among open-weight entries, and above base MiniMax H3 at 1277, fal's H3 Max, and Seedance 2.5. The gap over the base model it was post-trained from is 16 Elo, which is a real but narrow margin on a board where votes keep accumulating and positions move. Pika's press release does not publish a confidence interval or a vote count alongside the 1293, so the stability of seventh place is not something the launch materials let anyone assess.

The comparison wins are Pika's own, and the metrics are unnamed

Alongside the DesignArena number, Pika reports that H3 Flex won "the overwhelming majority of comparisons" against MiniMax H3, fal's H3 Max and HeyGen Video's quality model across all four video-quality metrics it tested, and that a separate human preference evaluation put Flex highest for human anatomy, complex composition, audio sync, visual style and camera control. Neither evaluation comes with the four metric names, the number of prompts, the number of raters, or a single percentage. The release then describes DesignArena's results as mirroring those internal evaluations, which is a claim about agreement between one disclosed number and two undisclosed ones.

Seven seconds is transformer compute, not the wait on the API

The speed claim is that a 15-second clip with audio at 768p can take as little as seven seconds, which Pika presents as faster than real time. Pika's own launch post qualifies this: that figure is transformer computation and not the full round trip of an API response. The price and speed claims are also relative rather than absolute, stated as up to 1.8 times faster than fal's H3 Max and up to 80 percent cheaper, with no per-second rate in the release itself.

H3 Flex is available now through the Pika API Club and as open weights, with support for up to 12 reference files per generation. It reaches the Pika creation product on November 2, which means the version most people will actually use is the hosted one, with the upscaler and the stems that the open release leaves out.

Sources: Pika (Business Wire), Pika, Pika on X, Hugging Face, GitHub.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.