Pika H3 Flex vs H3 Max: Weights, Speed, and Workflow

Pika H3 Flex is not a self-hosted edition of H3 Max. Pika released a six-step H3 adapter stack that can run locally or through Pika's hosted API. H3 Max is a separate fal post-training project, also sold through hosted endpoints. The useful comparison therefore has three paths: Local Flex, Hosted Flex, and Hosted H3 Max.

That distinction clears up two easy-to-misread launch claims. Flex generates natively at 480P or 768P; its hosted 1080P, 2K, and 4K outputs add an upsampler after a 768P generation. Pika's “seven seconds” for a 15-second 768P clip is controlled Transformer computation, not the time from API submission to a downloadable file.

What did Pika actually release?

Pika's official H3 Flex repository contains inference code and two post-trained adapters built on MiniMax H3:

• FL2VA covers text, first frame, last frame, and first-and-last-frame conditioning.
• Ref2VA accepts ordered image, video, and audio references.

Both adapters produce synchronized video and stereo audio with six generator evaluations. The local contract is specific: integer requests from 5 to 15 seconds, 24 fps, 32 kHz stereo audio, native 480P or 768P, and 11 aspect-ratio buckets. Its frame grid makes the files slightly longer than the request: five seconds becomes 124 frames, about 5.17 seconds, while 15 seconds becomes 362 frames, about 15.08 seconds.

Ref2VA accepts up to nine images, three videos, and three standalone audio files, with no more than 12 files in one request. At least one image or video is required. Each video or audio file must run 2–15 seconds; combined video duration and combined standalone-audio duration each stop at 15 seconds. Sound embedded in a reference video stays with that video and does not consume a separate audio slot.

“Up to 12 references” describes the input validator, not a guarantee that the largest possible bundle will fit a particular GPU. Pika explicitly says its verification table does not qualify every duration, aspect ratio, and reference combination.

H3 Max shares the MiniMax H3 ancestry but is not one of these adapters. fal describes H3 Max as its own post-trained model and H3 Max Turbo as its higher-throughput route. MiniMax also lists MiniMax-H3-Max in its direct V2 API. This site's ordinary generator calls fal H3 Max Turbo, not Pika H3 Flex and not MiniMax's direct Max endpoint.

The FastH3 versus H3 Max guide covers a different public H3 derivative. MiniMax H3 versus H3 Max explains the base-model split.

What can Local Flex actually run?

The current Python package is named pika-h3-pro. It requires Python 3.12, an NVIDIA GPU and compatible driver, FFmpeg, and compiled dependencies. Pika's hardware notes report representative tests on one H100 80GB per request; they do not declare H100 the only possible or minimum GPU.

The published numbers still show the scale of the capacity-planning problem:

• one request uses one GPU, and extra visible GPUs do not shard or accelerate it;
• representative five-second 480P sampling allocated roughly 66–69 GiB;
• the tested 15-second 768P case exceeded an H100 80GB without CPU offload;
• with offload, that case allocated about 29–32 GiB on the GPU and about 94 GiB of host memory including page cache;
• allow at least 211 GB for base components plus about 5.5 GB for both adapters, before dependencies and download staging.

The six-step claim also belongs to a particular model schedule, not a generic sampler setting. The release uses six denoising forwards with fixed knots, video and audio shifts, and fresh intermediate noise. Setting ordinary H3 to “six Euler steps” does not reproduce Flex.

Pika's launch post says a 15-second 768P clip can use seven seconds of Transformer (DiT) computation in a controlled benchmark. Its technical deep dive separately reports a five-second 768P text-to-video acceleration test on eight B300 GPUs. Neither number predicts a local H100 job or a hosted click-to-file wait. fal's timings.inference likewise describes backend inference rather than upload, queueing, encoding, storage, and download.

Local Flex is therefore a credible route for a team that needs to inspect or modify the adapters and has the infrastructure to qualify them. It is not a lightweight desktop download.

What stays on Pika's hosted side?

Pika sells the Flex family as a managed API, so using Flex does not require owning a GPU stack. Hosted Text-to-Video, Image-to-Video, and Reference-to-Video remove model provisioning while retaining Pika's request formats.

The downloadable generator stops at native 480P or 768P. Pika's hosted text and image endpoints can return 1080P, 2K, or 4K in one managed job, and its Draft to video endpoint can upgrade a selected draft. In both cases, Pika documents the higher-resolution path as a 768P generation followed by one upsampler pass. A truthful sample label would read:

Pika H3 Flex · native 768P · Pika upsampler · final 4K

That is a finishing workflow, not native 4K diffusion.

Audio-stem separation is another hosted extension. Pika can return Dialogue, Score, SFX, Foley, and Ambience as separate stems; the local adapter package produces a mixed stereo track.

Prompt enhancement needs a finer distinction than “paid add-on.” Pika says its Prompt Rewriter is automatically included at no extra charge in Pika-hosted Flex generations. A team running Flex itself can call the independent Prompt Rewriter API, currently listed at $0.02 per request.

Checked October 31, Pika API Club listed a $10 monthly membership that includes $10 in API credits each month. Its live Flex Text-to-Video pricing showed launch rates of $0.01 per successful output second at 480P and $0.02 at 768P, with regular rates displayed as $0.02 and $0.04. Reference-to-Video launch rates started at $0.03 per output second for 480P and $0.06 at 768P, and it also bills reference-video seconds; the displayed regular output rates were $0.06 and $0.12. Pika did not state an end date for the launch rates. Higher-resolution finishing and audio stems use their own rates.

For comparison, fal's H3 Max Turbo promotion shows $0.026 per second at 768P through October 31. Its published schedule moves to $0.028 per second from November 1 through November 15, then to the $0.04 standard rate on November 16. A five-second 768P API job is therefore $0.13 on October 31, $0.14 during the November promotion, or $0.20 at the standard rate. These are dated provider prices, not this site's retail credits. An ordinary job here costs three credits per second at 480P or five credits per second at 768P, and the Generate button shows the total before submission.

Which license governs Local Flex?

Pika licenses its inference and loader code under Apache 2.0. The Flex adapters are derivatives of MiniMax H3 and remain governed by the MiniMax H3 Community License Agreement.

The current agreement excludes the United States, European Union, United Kingdom, and South Korea from its Applicable Territory. It also requires separate prior authorization for commercial products or services above $20 million in annual revenue and carries distribution, attribution, user-term, and safety obligations.

That split matters: Apache-licensed code does not make the accompanying model derivatives unrestricted “open source.” Read the live agreement for the deployment and territory you plan. Pika's hosted API and hosted H3 Max instead place the model server behind provider service terms, so their legal and operational contracts are different from distributing or serving the local adapters yourself.

Local Flex, Hosted Flex, or H3 Max?

Two requirements make the decision quickly. If you must modify or operate the model, start with Local Flex and clear its hardware and license constraints first. If you want a managed service, compare Hosted Flex with the exact H3 Max route on the inputs you need: plain text, first/last frames, or a mixed reference pack. “No GPU provisioning” is not unique to H3 Max; both hosted paths provide it.

H3 Max can also handle reference-driven work through fal's separate Reference-to-Video endpoint, but this site's ordinary Turbo form does not expose that request body. The H3 Max Reference-to-Video guide lists its inputs and billing boundaries.

Pika H3 Flex is interesting because Pika released a fast, reference-rich H3 adapter stack and offers a managed version of it. H3 Max is a different hosted model contract. Compare the path you can actually operate, the inputs your shot needs, and the generation stage behind the advertised resolution; the family name alone cannot answer any of those questions.