Pika Drops H3 Flex, an Open Video Model Generating Audio 1.8x Faster

Pika has released Pika H3 Flex, an open-weight video model that generates synchronized video and stereo audio with six denoising steps. The model is post-trained from MiniMax H3, with adapter weights on Hugging Face, inference code in the GitHub repository, and integrations for ComfyUI and diffusers.

At the time of Pika’s announcement, H3 Flex ranked seventh on DesignArena’s text-to-video leaderboard and led its open-weight entries. Pika also reports that its hosted API can produce a 15-second clip with audio in as little as 7 seconds, meaning the output runs longer than the generation process.

Two adapters cover five workflows

Pika H3 Flex keeps MiniMax H3 as its base and adds two task-specific adapters. FL2VA handles prompts and endpoint frames, while Ref2VA conditions generation on image, video, and audio references.

The model accepts up to 12 reference inputs, including sketch, voice sample, image, sound, video, and moodboard. That mix supports character, scene, motion, and soundtrack references within one request.

Pika says the hosted upscaling stage preserves the draft’s motion and audio while increasing resolution. Developers who need direct local inference are limited to the two native resolutions in the open release.

A custom six-step sampler

H3 Flex uses six denoising model evaluations across seven timestep boundaries: 1000/750/500/250/100/40/0. A denoising evaluation is one pass that converts a noisier intermediate representation into a cleaner one, so reducing the pass count directly lowers generation latency.

The adapters also use video shift 12, audio shift 3, and fresh noise at intermediate stages. The repository warns that a generic six-step Euler sampler follows a different process, because the released schedule is tied to the adapter training.

Pika attributes the faster-than-real-time result to this schedule. Its matched comparison reports roughly 1.8 times faster generation than MiniMax H3Max. Actual local latency will vary with GPU, resolution, clip length, and CPU offloading.

Audio arrives ready to mix

Local inference produces a mixed stereo track alongside the video. The hosted API can also return five separate .wav stems: Dialogue, Score, SFX, Foley, and Ambience.

Separate stems let developers send generated material into a nonlinear video editor or digital audio workstation without first attempting source separation. Pika charges an additional $0.001 per generated second for stem export.

Local inference needs substantial hardware

The README recommends Python 3.12 and an NVIDIA GPU, and Pika tested representative requests on a single H100 with 80 GB of memory. Downloading both tasks requires about 215.5 GB, comprising roughly 210 GB of base files and 5.5 GB of adapters.

• Each request runs on one GPU. Single-clip sharding across multiple GPUs is unavailable.
• The --cpu-offload option reduces VRAM use by moving data through host memory, increasing RAM demand and transfer time.
• Output lengths follow a 17n+5 frame grid at 24 fps.
• A nominal 5-second request produces 124 frames, or about 5.17 seconds.
• A nominal 15-second request produces 362 frames, or about 15.08 seconds.

The repository’s command-line interface exposes the task, model revision, resolution, aspect ratio, duration, and seed:

pika-h3-pro --task t2va --prompt-file examples/prompts/t2va.txt \
  --model-repo "$MODEL_REPO" --revision "$MODEL_REVISION" \
  --resolution 480p --aspect 16:9 --duration 5 --seed 42 \
  --output outputs/sample.mp4

Hosted pricing trades hardware for resolution

Launch pricing on the Pika API starts at $0.01 per output-video second for 480p and reaches $0.08 per second for 4K. Pika describes those rates as roughly half its regular pricing.

At the disclosed endpoint rates, 15 billed seconds cost $0.15 at 480p or $1.20 at 4K before stems and reference-video charges. Pika also announced a launch promotion of 15 free daily generations for API Club members during the first two weeks.

The leaderboard lead is narrow

DesignArena summarizes head-to-head user preferences as Elo ratings. H3 Flex recorded an Elo score of 1293 and ranked seventh overall when Pika announced the release.

The 16-point lead over MiniMax H3 indicates that Pika’s post-training improved the base model’s standing on this evaluation. Elo scores remain relative and change as votes accumulate, so the figures capture the board at announcement time.

A reusable adapter strategy

Pika’s release shows how a team can specialize an existing video model, outperform its base on a preference leaderboard, and distribute the resulting adapters with inference code. The six-step schedule is especially relevant to developers because it connects post-training choices directly to lower inference latency.

The release also offers three deployment paths: local generation at 480p or 768p, hosted generation at higher resolutions, and node-based workflows through the live ComfyUI integration. The adapter weights are available on Hugging Face.

The MiniMax H3 Community License governs use of the released adapters, so teams should review its terms before commercial deployment. Local operation also carries a high storage and GPU cost, making the hosted API the more accessible route for developers without server-class NVIDIA hardware.