Pika H3 Flex Ships Open Weights for 768p. 4K, Five Audio Stems, and Fast Serving Stay on Pika's Hosted Stack.
Pika released H3 Flex on October 29: MiniMax H3 post-trained to generate in 6 steps instead of 50, with up to 12 reference inputs and synchronized stereo audio, with generator adapters and inference code on Hugging Face and GitHub under MiniMax's H3 Community License. Local generation is 480p and 768p; the upsampler to 1080p, 2K, and 4K, the five separated audio stems, prompt enhancement, and accelerated serving are Pika-hosted. Pika's seven-second figure for a 15-second 768p clip is transformer compute, not end-to-end latency, by its own footnote. The creation product lands November 2; there is no API price on the launch post. We read the launch post and the repositories.
By T.W. Ghost
Pika released H3 Flex today, October 29, 2026. It is MiniMax's H3 video model post-trained to generate in 6 denoising steps instead of 50, with up to 12 reference inputs and synchronized stereo audio. The generator adapters and inference code are on Hugging Face and GitHub, which makes this an open-weight release, and the license on the adapters is MiniMax's H3 Community License, not Apache. The creation product inside Pika arrives November 2.
The split to understand before you download anything: what runs locally generates 480p and 768p. The upsampler that takes output to 1080p, 2K, and 4K, the separation of audio into five stems, Dialogue, Score, SFX, Foley, and Ambience, prompt enhancement, and the accelerated serving path are all on Pika's hosted stack. The headline speed, a 15-second 768p clip in as little as seven seconds, carries Pika's own footnote: that is transformer compute, not end-to-end API latency. We read the launch post, the Hugging Face repository, and the GitHub repository. Here is what held up.
Open and Hosted, Side by Side
| Runs from the open release | Pika-hosted only | |
|---|---|---|
| Generation | 480p and 768p | |
| Resolution above 768p | Upsampler to 1080p, 2K, 4K | |
| Audio | Synchronized stereo | Five separated stems via API |
| References | Up to 12 | |
| Prompt enhancement | Hosted | |
| Fast serving | Hosted | |
| Steps | 6 (H3 base: 50) | |
| License | MiniMax H3 Community License on the adapters; confirm the LICENSE file | |
| API price | Not on the launch post |
Open weights are not the product. The thing Pika sells on November 2 is the hosted stack. The thing on Hugging Face is the generator at two resolutions, and that is a fair description of what "open" means here.
Seven seconds is a DiT number. Pika says so in its own footnote. Wall-clock time to a finished clip includes everything the footnote excludes.
Read the license before calling it open. The code may be Apache; the adapters carry MiniMax's community license, which is a different document with its own terms. Pika states the license; we did not independently read the LICENSE file this pass.
The Benchmarks, Such as They Are
Pika says H3 Flex ranked first among open-weight models and seventh overall on DesignArena's text-to-video board at launch. That is a launch snapshot cited by the vendor, the same board we cautioned about re-pulling last week. There is no Artificial Analysis video arena entry we could find, and no independent quality comparison to the MiniMax H3 base it was trained from.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 29; post-trained MiniMax H3; 6 steps vs 50 | Verified (Pika launch post) |
| Up to 12 references; synchronized stereo audio; five stems via API | Verified (launch post) |
| Native 480p and 768p; upsampler to 1080p, 2K, 4K hosted | Verified (launch post) |
| 15-second 768p clip in about 7 seconds of DiT compute | Vendor, with Pika's own not-end-to-end footnote |
| DesignArena first open, seventh overall | Vendor citation of a launch snapshot |
| Weights and code on Hugging Face and GitHub; product Nov 2 | Verified (launch post; repositories) |
| MiniMax H3 Community License on the adapters | Verified as Pika's statement; LICENSE file not read here |
| API price | Unverified (not on the launch post) |
Who Should Use It
Download H3 Flex if 768p with synchronized audio at 6 steps is useful on your own hardware and MiniMax's community license fits your use; that is a real release.
Skip the "open 4K video model" framing, because 4K is the hosted upsampler, and skip quoting seven seconds without the word "compute" next to it.
The frontier is elsewhere: DeepMind's Genie 4 is the week's other video news and it has no weights, no price, and no date; the hosted leader on the arenas is still Seedance 2.5. Our model comparison has the current standings on the language models side by side.
Sources
*Not sure whether a 768p open model or a hosted 4K pipeline fits your video work? Take the free 2-minute quiz and get matched. Then read about the world model that promises everything and ships a waitlist: Genie 4.*