Pika releases H3 Flex, an open-weight video model with sound

Pika has released H3 Flex, a post-trained version of MiniMax’s H3 video model that generates video with synchronized audio in six sampling steps and accepts up to 12 reference inputs. The model page on Hugging Face offers weights for local inference at native 480p and 768p, while Pika’s hosted API extends the same model to 1080p, 2K and 4K output.

This is a step beyond the two earlier Unicorn API reports in this thread. The first covered Pika’s API documentation for the base MiniMax H3 model, which offers 768P and 2K clips of 4 to 15 seconds. The second covered Pruna’s listing of three H3-based video models. What is new now is a Pika-branded model of its own, with downloadable weights, aimed at what Pika calls “professional iteration”.

What Pika says H3 Flex is built for

Pika describes H3 Flex as “a post-trained H3 model, designed to give professional creators the flexibility and speed to explore with precision”. Post-training means further training of an existing model to change its behaviour, in this case starting from MiniMax H3. Pika says the model is “optimized for greater control” and that shots can be directed through both references and text prompts.

The model page lists the uses Pika has in mind: multi-shot storytelling, motion graphics, product films and reference-guided animation. Its sample prompts show what that looks like in practice. One asks for a 15-second table-tennis rally with a different camera move timed to each three-second block, from a top-down shot to a crane down, whip pans, a 180-degree orbit and a low-angle dolly. The prompt also specifies sound effects and a drum-and-bass groove. Another describes a single continuous shot around an apartment tower, with second-by-second camera directions and a request for natural location sound only, with no music or voices.

These are Pika’s own showcase prompts. The sources do not include independent clips or tests showing how closely the model follows such detailed instructions.

What H3 Flex changes compared with base MiniMax H3

The clearest headline figure is speed: Pika says H3 Flex generates video with synchronized audio in six sampling steps. A sampling step is one pass in which the model refines its output, so fewer steps generally means less compute per clip. The base H3 model card in the sources does not state a step count, so the size of any gain cannot be measured from the sources.

On references, H3 Flex supports up to 12 inputs. That matches the cap in base H3’s “omni-reference” mode, which allows up to 9 images, up to 3 video clips and up to 3 audio clips, with a maximum of 12 files across all types. The Flex page does not break its 12-input limit down by type.

Resolution is split between local and hosted use. Base H3 generates at 768p by default and reaches 2K through a second module, H3-Regenerate-2K, that feeds the 768p result and the original context back into the model. H3 Flex runs locally at native 480p or 768p, and Pika’s hosted API adds 1080p, 2K and 4K. The sources do not say how Pika produces its 4K output.

Audio also differs by route. Base H3 outputs 32 kHz stereo audio. Local H3 Flex runs produce mixed stereo audio, while the hosted API adds audio stems, meaning separate audio tracks instead of a single mix. The page does not say which tracks the stems contain.

Spec	Pika H3 Flex	Base MiniMax H3
Sampling steps	Six	Not stated in sources
Reference inputs	Up to 12	Up to 12 files (≤9 images, ≤3 videos, ≤3 audio clips)
Local resolution	Native 480p / 768p	768p default; 2K via H3-Regenerate-2K
Hosted resolution	1080p, 2K, 4K via Pika API	768P or 2K on Pika’s API
Audio	Mixed stereo locally; stems via Pika API	32 kHz stereo
Clip length	Not stated on Flex page	4–15 seconds
Frame rate	Not stated on Flex page	24 FPS

Open weights, hosted API, and what stays closed

The H3 Flex page links model weights, a quickstart, a prompt guide and Pika API extensions, and states that “local open-weight inference supports native 480p / 768p video with mixed stereo audio”. In other words, anyone can download the model and run it on their own hardware, but the higher resolutions, audio stems and prompt enhancement are features of Pika’s hosted service.

The page text reviewed does not state a licence for H3 Flex. A separate MiniMax H3 model card on Hugging Face, published under the account “lowhigh”, refers to an open-source release of H3 governed by the MiniMax H3 Community License, including obligations on lawful use and use restrictions. Builders planning commercial use should check the licence attached to the H3 Flex weights directly.

That H3 card also explains what the base open release leaves out. MiniMax H3 relies on a preprocessing system, H3-Context-IR, that interprets text, images, audio and reference videos and turns them into a structured input for the generator. Because it depends on “multiple hosted models and services”, it is “not included in this open-source release”; instead, an API and prompting guidance are offered so developers can build their own preprocessing. The card calls this step “critical to the quality of the final output”. Pika’s hosted API offers “prompt enhancement” for H3 Flex, but the sources do not say whether this is the same system or how local runs compare without it.

Two further technical notes from the base H3 card matter for anyone running weights locally. Its text encoder uses the full pretrained weights of Qwen3-VL-32B, a 32-billion-parameter vision-language model. And although H3 supports sparse attention, a technique that cuts compute on long sequences, the initial open release offers inference with full attention only, with sparse attention promised in a later update. The sources do not say whether either point applies unchanged to H3 Flex.

What the evidence shows, and what it does not

The evidence for H3 Flex comes almost entirely from Pika’s own model page. Every capability above, from six-step generation to 4K API output, is Pika’s description of its own model. No source contains an independent benchmark, leaderboard result or third-party test of H3 Flex.

Independent hands-on material exists for the base model only. Atlas Cloud, a model hosting provider, lists blog posts testing MiniMax H3, including one that checks whether its stereo audio is genuinely stereo, a music-video experiment, and a comparison of Seedance 2.5 with MiniMax H3. Only the post titles are visible in the source, and none concerns H3 Flex.

A community GitHub repository for MiniMax H3 offers ComfyUI workflows, a prompt builder, a VRAM calculator covering consumer and data-centre GPUs with quantization options, and “benchmark comparisons” against other video models. The repository states that it is “not affiliated with, sponsored by, or endorsed by MiniMax Inc.”, and its contents are not evidence of model quality.

What H3 Flex means for builders

For teams building video features, H3 Flex offers a choice between two routes from one model family. Local open weights suit prototyping, look-development and pipelines that need to keep media in-house, within the 480p and 768p limits. The hosted API suits final renders that need 1080p to 4K or separate audio stems.

The six-step claim matters most for iteration-heavy work, where a creator generates many variants before choosing one. If it holds up in testing, it could make local drafting more practical, but the sources give no timing figures or hardware requirements for H3 Flex.

Several limits apply. Pika’s API documentation lists what base H3 is not suited to: frame-exact rotoscoping, feature-length renders, pixel-perfect compositing and precise on-screen text. The sources do not say whether H3 Flex changes any of these. The base H3 system also applies automated moderation to submitted text, images, videos and enhanced prompts, and its card warns that filters “cannot eliminate false positives or false negatives”.

Cost is an open question. Pika’s MiniMax pricing page lists per-second rates for base H3 across text-, image- and reference-to-video, but shows no separate entry for H3 Flex. Builders cannot yet compare the price of a Flex render with a standard H3 render from the sources.

Open questions the sources leave

• Licence: the terms attached to the H3 Flex weights are not stated in the page text reviewed.
• Clip length and frame rate: the Flex page does not state them; base H3 runs 4 to 15 seconds at 24 FPS.
• Languages: base H3 reports stable dialogue support in 11 languages; the Flex page does not say whether post-training affects this.
• Quality trade-off: no source shows whether six-step generation costs detail or motion quality compared with base H3.
• Hardware: no official minimum GPU requirement for local H3 Flex appears in the sources.

What to watch next

The next useful evidence will be independent: third-party tests and leaderboard entries that compare H3 Flex with base H3 and rival video models on prompt adherence, motion and audio sync. Also worth watching are Pika’s pricing for H3 Flex on its API, a clear licence statement on the weights, and whether MiniMax’s promised sparse-attention update for the open H3 release reaches the Flex version and lowers local compute needs.