Pika H3 Flex releases weights with regional license exclusions

Pika released the weights for Pika H3 Flex on October 29. The post-trained MiniMax H3 generator produces synchronized video and audio locally at 480p or 768p. Higher-resolution finishing and separate audio tracks are hosted features.

Its published MiniMax H3 Community License excludes South Korea, the United States, the United Kingdom and the European Union from the permitted territory. It directs prospective deployers in those regions to contact MiniMax for a separate license.

The model card specifies 5–15-second requests, 24 frames per second and up to 12 references overall. The individual ceilings are nine images, three videos and three audio clips, subject to that combined limit. Reference mode requires an image or video; combined video duration and combined audio duration are each capped at 15 seconds. Individual clips must be 2–15 seconds. Output audio is 32 kHz stereo. Maximum accepted reference combinations are not guaranteed to fit GPU memory.

ComfyUI supplies text, reference-image and first/last-frame workflows. Endpoint images anchor the opening and closing shots, while a prompt describes the movement between them. Its published testing notes report imperfect camera direction and transition timing, plus unreliable fine contact distances between hands and objects.

Six steps and the hardware behind them

Pika describes three training stages: supervised audio-video learning, distillation from 50 denoising steps to six, and reward and preference refinement. In a separate optimization test, full inference for five-second 768p video with audio fell from 3.409 to 2.129 seconds on eight B300 GPUs. This company measurement kept the checkpoint and six-step schedule fixed.

The downloadable package uses one GPU per request. Pika tested representative requests on an H100 80GB: 480p, five-second sampling allocated about 66–69 GiB. CPU offload supports 768p, 15-second sampling. Those figures exclude separate encoder and decoder requirements and are measurements, not minimum specifications.

Base components require at least 211 GB of storage, plus approximately 5.5 GB for both adapters and further room for dependencies and download staging. ComfyUI says existing H3 users can reuse their base files and add the two adapters. H200 and B300 remain outside the downloadable package’s qualified configurations. Example Python code has an Apache 2.0 license, which does not cover weights or example media.

The hosted API can return five WAV tracks: dialogue, score, effects, foley and ambience, allowing separate sound adjustments during editing. Stems cost $0.001 per video second at regular rates.

Hosted finishing upscales a 768p draft to 1080p, 2K or 4K. The downloadable release contains generator adapters and inference code; Pika operates the upsampler, stem separation and accelerated serving. Its seven-second claim for a 15-second clip measures transformer computation in a controlled benchmark. Input processing, finishing and service overhead affect total API completion time.

Current text, image and keyframe pricing lists 768p at $0.02 per output second during launch, versus $0.04 regularly: $0.30 or $0.60 for 15 seconds. Video-reference generation uses a separate rate charged on input plus output duration. The product page dates API Club’s 15 free daily generations from October 29 through November 5.

The license also requires prior written authorization when commercial products and services exceed $20 million in yearly revenue, and display of the MiniMax H3 name on commercial interfaces.