SSyncDev

Pika H3 Flex

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Pika H3 Flex is Pika's open LoRA post-training of MiniMax H3: six-step sampling, up to 15 s clips with 32 kHz stereo audio at 480p or 768p, 24 fps.

Key specifications

Specification score
24.9
License
MiniMax H3 Community License (adapters as model derivatives); example code Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights are free to self-host. Hosted access runs through the Pika API; no price was published at launch, so check your own account.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input
  • Video input
  • Audio input

Produces

  • Video output
  • Audio output

Capabilities

VisionAudioOpen weights

What is Pika H3 Flex?

Pika H3 Flex is not a new foundation model, and that matters before you pull roughly 210 GB of files. It is Pika's post-training of MiniMax H3, released on 29 October 2026 as two LoRA adapters plus inference code. You bring, or download, the base; the adapters change how it samples.

The headline change is speed. Generation takes six steps instead of 50.

Two adapters, two jobs

  • fl2va covers text-to-video, first-frame, last-frame, and first-and-last-frame conditioning.
  • ref2va takes a prompt plus ordered references: up to 9 images, 3 videos and 3 audio clips, 12 files at most, with at least one image or video. Each video or audio reference must run 2 to 15 seconds.

Both emit synchronised stereo sound with the picture, so dialogue, effects and music arrive in the same pass.

The output envelope

Property Value
Repository Pika-Labs/Pika-H3-Flex
Base MiniMax H3 (listed at 33B parameters)
Native resolution 480p or 768p, 11 aspect ratios from 21:9 to 9:21
Frame rate 24 fps
Duration 5-15 s, whole-second requests
Audio 32 kHz stereo, muxed into MP4 (H.264 + AAC)
Prompt length 1-8,000 characters
Sampling Six steps with the adapter's own schedule
Licence MiniMax H3 Community License (adapters); example code Apache-2.0

Durations snap to a frame grid, and this is where two published figures disagree. The Diffusers snippet uses 121 frames for five seconds, but the grid itself maps a 5 s request to 124 frames, about 5.17 s. If you set frame counts by hand, use grid values, or let the CLI round for you, then check the real length of what comes out.

A note on timing: Pika says a 15-second 768p clip needs about 7 s of DiT compute in its own benchmark. That excludes decoding, audio and any CPU offload, so it is not end-to-end latency. Generic six-step Euler sampling also does not reproduce the result; the adapter's schedule is required.

Running it yourself

Representative 480p and 768p runs pass on a single H100 80GB. Bigger requests may need --cpu-offload and plenty of host RAM. In Diffusers you load a modular pipeline with workflow set to fl2va or ref2va, keep the two VAEs in FP32 and everything else in BF16, and load the matching adapter explicitly, because the bundled base components are not Pika-adapted.

The adapters are model derivatives of MiniMax H3, so its community licence and a NOTICE file travel with them. The Apache-2.0 example code covers neither weights nor sample media.

Local versus hosted

Some features stay on Pika's API: upscaling to 1080p, 2K or 4K from a draft, five-stem audio separation and prompt enhancement. No API price was published at launch.

Launch dating varies slightly; some trackers file it a week later than 29 October. The weights are the same either way.

To judge it, render the same twenty prompts on the base model at 50 steps and on the adapters at six. Compare motion coherence and lip sync, and time the full pipeline rather than the DiT pass. For other open and hosted options see text-to-video models, or the rest of the MiniMax line.

How to evaluate Pika H3 Flex

Compare the model on the work you need to complete.

Tasks and inputs

Pika H3 Flex is listed for Image-to-video, Text-to-video, Audio-Driven Video. It accepts text, image, video, audio and produces video, audio. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.

Choosing a deployment

The weights are downloadable under MiniMax H3 Community License (adapters as model derivatives); example code Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

No. It is a pair of LoRA adapters plus inference code on top of MiniMax H3, which you also need to download.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Google

Gemini Omni 1.2 Flash

Gemini Omni 1.2 Flash is Google's conversational video generation and editing model, making 3-15 second clips with native audio from 360p to 4K.

131.07K context

heygen

HeyGen Video 1

HeyGen Video 1 generates 5 to 30 second MP4 clips with audio from text, images or reference media, up to 2K at 24 fps. Built on MiniMax H3.

— context

Meta

Muse Video

Muse Video is Meta’s previewed video-generation model with native audio; a public Meta Model API route and price have not been established.

— context

Abiray

Abiray/LTX-2.3-22B-DISTILLED-1.1-GGUF

Abiray/LTX-2.3-22B-DISTILLED-1.1-GGUF is a source-linked Hugging Face repository assigned to Text-to-video. Its recorded task interface accepts text and produces video.

— context