On October 29, 2026, Pika released H3 Flex: adapter weights and inference code for a video model post-trained on MiniMax H3, generating video with synchronized stereo audio in six sampling steps. The word "open" traveled fast. But if you download the package expecting the full feature list from the announcement, you are in for a confusing afternoon—because Pika has drawn a deliberate line between what the weights do on your own GPU and what remains a hosted, metered service. That line is the whole story here.

What the local package actually gives you

The released package is genuinely usable, not a token gesture. You get two task adapters—one for text and endpoint-frame conditioning, one for multimodal references—and a pipeline that produces MP4 files with H.264 video and AAC stereo audio baked in. Local generation runs at native 480p or 768p, across eleven aspect ratios, at 24 frames per second, with 32 kHz stereo audio mixed into the picture. Requested durations are 5 to 15 seconds, though the model works on a frame grid, so a "5-second" request actually yields about 5.17 seconds of footage.

The reference system is the interesting part for professional work: up to twelve reference files per generation—nine images, three videos, three standalone audio clips—so you can steer characters, motion, and sound with material you supply. This is the "Flex" pitch: iteration control for people making product films and multi-shot sequences, not one-prompt slot machines.

What stays in the cloud

Three headline capabilities are hosted-only, and the repository says so plainly. Higher-resolution output—1080p, 2K, and 4K—comes from the API's draft-to-video endpoint, which takes your draft and upscales it server-side. Separate audio stems, for pulling apart the generated soundtrack into editable tracks, are an API feature. So is prompt enhancement, the rewriter that polishes your prompt before generation. The pricing page lists the meters: text-to-video and image-to-video from $0.01 per second, reference-guided generation from $0.03 per second, prompt rewriting at $0.02 per request, audio stems at $0.001 per second.

So if you saw a glossy 4K clip in the launch material and want to reproduce it on your own machine: you cannot, and that is by design. The local package drafts at 480p/768p; the 4K finish is a service with a per-second price. The creative-product entrance Pika announced for November 2 sits on the hosted side as well.

The license split that trips people up

Here is a detail worth reading before you build anything commercial. The inference and loader code is Apache-2.0—permissive, do what you like. The adapter weights are not. They are distributed as Model Derivatives of MiniMax H3 under the MiniMax H3 Community License Agreement, which carries its own terms and a required notice. Code license and model license are separate documents answering separate questions, and "the repo says Apache" does not describe the weights. Read the community license before the weights go anywhere near a product.

The hardware bill

Open-weight does not mean laptop-friendly. The published test configuration is one H100 with 80 GB of memory per request; H200 and B300 were untested at release. Downloading both task adapters pulls about 210 GB of pinned base files plus 5.5 GB of adapters. One request runs on one GPU—the package does not shard a generation across cards—and there is an optional CPU-offload mode that trades host RAM and transfer time for lower VRAM pressure. Pika's own advice for qualifying your setup: start at 480p, five seconds.

About the speed claims

The announcement's eye-catching number—seven seconds of compute for a 15-second video—measures model compute under a specific benchmark setup. It is not the time from pressing enter to watching your clip, which also carries encoding, decoding, downloads, and whatever queue your hardware is actually running. Treat it as a laboratory figure for the sampler, not a promise about your render farm. The same caution applies to the showcase videos: the repository itself notes they are creative demonstrations, not release-validation renders, and that exact reproduction with the released package is not guaranteed. Even the six-step sampling schedule is specific—generic six-step Euler sampling is not equivalent, and results can drift across attention kernels, devices, and library versions even with identical seeds.

Who this is for

If you are a video-curious developer with access to serious GPU capacity and you want a controllable, audio-synchronized generation pipeline you can actually inspect and modify, this is one of the most complete open releases in the category—workflow, prompt guide, hardware tables and all. If you wanted free 4K video on a gaming PC, the honest answer is that neither word of that sentence survives contact with the documentation. The weights are open; the resolution, the stems, and the polish are a business.

We have not run the package ourselves; everything above comes from Pika's release documentation and API listings.