Pika-H3-Flex Audio-Visual Video Generation Model: File List

Pika-H3-Flex Audio-Visual Video Generation Model: File List

At a Glance

Item Value
Repository Pika-Labs/Pika-H3-Flex
Family guide MiniMax-H3 guide (14 articles)
Publisher guide MiniMax: models and licenses
Published 2026-10-30
License minimax-h3-community-license-agreement
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Pika-Labs has released “Pika H3 Flex", a model adapter capable of generating videos with synchronized stereo audio. This model accepts text, images, and various reference media as inputs to simultaneously generate video and audio.

It adopts MiniMaxAI’s “MiniMax-H3" as its base model, optimized to generate videos with synchronized stereo audio in a short time of just 6 sampling steps. For local inference, it supports resolutions of 480p and 768p, and accommodates multimodal conditioning using up to 12 reference files.

Specifications

The specifications of Pika H3 Flex and the base model MiniMax-H3 as stated in the public documentation are as follows:

  • Model Configuration & Adapter Format: Provided as FP32 precision LoRA (rank 128 / alpha 128). Depending on the task, two types of adapters are available: “FL2VA" (for text and start/end frame control, base trunk: transformer) and “Ref2VA" (for multimodal references, base trunk: transformer_ref).
  • Base Model Architecture: Based on MiniMax-H3. MiniMax-H3’s Omni-Transformer is a 33B parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches. The encoder uses Qwen3-VL-32B (using hidden states from the 50th layer), the visual VAE uses H3-VisualVAE (spatial 16x, temporal 4x compression, effective spatial downsampling 32x), and the audio VAE uses H3-AudioVAE with 32kHz stereo support that processes left and right channels independently.
  • Output Resolution: Native resolutions are 480p and 768p. It supports all 11 aspect ratio buckets (21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16, 9:21).
  • Output Framerate: 24 fps.
  • Output Duration (Video Length): Supports integer second specifications from 5 seconds to 15 seconds. The actual generated length is rounded up to the model’s frame grid, resulting in 124 frames (approx. 5.17 seconds) for a 5-second specification, and 362 frames (approx. 15.08 seconds) for a 15-second specification.
  • Output Audio: 32 kHz stereo audio mixed into the video output.
  • Output File Format: MP4 format including H.264 video and AAC audio, along with a JSON file of generation metadata.
  • Prompt Input Specifications: Supports direct text input or UTF-8 prompt files, with a character limit of 1 to 8,000 characters.
  • Step Count & Sampling: Operates with 6-step inference using a sampling schedule dedicated to the public adapter (noted as not identical to standard 6-step Euler sampling).
  • Generation Modes and Reference Limits:
  • FL2VA: Supports text-to-video-audio generation without input media (t2va), start frame specification (i2va), end frame specification (l2va), and both start and end frame specification (fl2va). – Ref2VA (ref2va): Accepts up to 12 reference files. This includes a maximum of 9 images, up to 3 video clips, and up to 3 independent audio clips (at least one image or video is mandatory). Videos and independent audio are 2 to 15 seconds per clip, with total duration limited to a maximum of 15 seconds each for video and audio.
  • License Terms: Governed by the MiniMax H3 Community License Agreement (minimax-h3-community-license-agreement).

Performance and Quality

Although the public repository does not list objective benchmark test measurements, the following details regarding the model’s output quality and operational characteristics have been published.

The model card showcases examples such as camera work control, cinematic storytelling, graphic design, and reference-guided animation. However, it notes that these are creative showcases demonstrating capabilities rather than formal rendering results or throughput benchmark measurements for performance verification.

Additionally, according to published materials on the original MiniMax-H3 model, MiniMax-H3 was designed from the pre-training stage as a multimodal foundational model that integrates and understands text, images, videos, and audio. It is reported to excel in following complex multimodal instructions and stably supports dialogue audio generation across a total of 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Pika H3 Flex inherits the high expressiveness of these base models while enabling simultaneous video and stereo audio generation in a fast inference cycle of just 6 steps, thanks to a dedicated sampling schedule.

Strengths and Use Cases

Pika H3 Flex is designed for professional iteration workflows where professional creators can quickly and precisely experiment. Because it can batch-generate video and synchronized stereo audio in an extremely short sampling process of just 6 steps, it is well-suited for accelerating the exploration of production processes and draft creation.

In public materials, specific use cases well-suited for the model include:

  • Multi-shot storytelling: Combining detailed scene instructions via prompts and camera work specifications to build video with cinematic direction and narrative progression.
  • Motion graphics and graphic design: Suited for generating animated design expressions and dynamic graphic video based on text instructions.
  • Product films (Product promotional video): Utilizing precise prompt instructions to control product textures and camera orbiting to quickly create drafts for commercial use.
  • Reference-guided animation: Leveraging reference media to generate animations while maintaining consistency in characters, specific art styles, and motion.

Supporting these use cases is flexible input control via two types of task adapters (FL2VA and Ref2VA). With FL2VA, users can not only specify just the start frame to generate subsequent progression, but also specify an end frame to build video leading up to it, or provide both start and end frames to interpolate and connect between them. This allows for precise direction of scene transitions and landing on intended compositions.

Furthermore, Ref2VA accepts exceptionally rich multimodal references: up to 9 images, up to 3 videos, and up to 3 independent audio tracks (up to 12 files in total). A major strength is the ability to carry out advanced directional instructions locally, specifying the appearance of a specific character with multiple images while combining sample motion videos and background music or ambient sound.

How It Differs from Similar Models

Previously covered derivative models based on MiniMax-H3 on this site include FP8 and NVFP4 Versions of Audio-Visual Video Generation Model “FastH3 V2" Released and Audio-Visual Video Generation Model “FastH3 Trim" 8-Step Version Released.

These models by the FastVideo team primarily targeted consumer-level GPU environments, focusing on lightweight optimizations such as weight quantization (FP8 and NVFP4) and the removal of transformer blocks (reducing 50 blocks down to 42 by removing 8), completing 8-step inference within a single model.

In contrast, the newly introduced Pika H3 Flex does not adopt an approach of shrinking and quantizing the model itself for distribution; instead, it is designed as a LoRA adapter applied to the base model MiniMax-H3. The key difference is that while compressing the inference steps even further to a short 6 steps, it provides two dedicated adapters—FL2VA (start/end frame control) and Ref2VA (up to 12 multimodal references)—enabling flexible direction without sacrificing expressiveness or control performance. While the FastH3 series prioritizes lightweight execution matching local resources, Pika H3 Flex aims to balance professional-grade precise direction leveraging multimodal reference materials with high-speed generation.

Distributed Files

Weight files published in Pika-Labs/Pika-H3-Flex, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
audio_vae/diffusion_pytorch_model.safetensors 605MB
comfy/loras/pika_h3_flex_fl2va_comfy_bf16.safetensors 1.96GB
comfy/loras/pika_h3_flex_ref2va_comfy_bf16.safetensors 1.96GB
fl2va/adapter.safetensors 2.77GB
ref2va/adapter.safetensors 2.77GB
text_encoder/model-*-of-00002.safetensors 66.71GB (2 split files)
transformer/diffusion_pytorch_model-*-of-00007.safetensors 66.28GB (7 split files)
transformer_ref/diffusion_pytorch_model-*-of-00007.safetensors 66.28GB (7 split files)
vae/diffusion_pytorch_model-*-of-00002.safetensors 10.42GB (2 split files)

How to Get It

Pika H3 Flex is available on Hugging Face and can be downloaded by anyone without special usage requests (gating).

  • Distribution Format: Weights are distributed in safetensors format. Official FP32 LoRA (rank 128 / alpha 128) weights for inference are stored under fl2va/ and ref2va/, while LoRA weights adjusted in BF16 format for ComfyUI are included in the comfy/loras/ directory. Additionally, Diffusers-format MiniMax-H3 base components necessary for operation (such as the Transformer backbone, Qwen3-VL-32B-based text encoder, visual VAE, and audio VAE) are bundled within the repository.
  • Source Repository: Available from Hugging Face at Pika-Labs/Pika-H3-Flex.
  • Inference via Official CLI: You can use it by cloning the inference repository published on GitHub (https://github.com/Pika-Labs/Pika-H3-Flex) and following the environment setup instructions for pip, Conda, or Docker. Running the provided pika-h3-pro command automatically downloads the required base components and selected adapter. The repository documentation suggests running it on an H100 80GB GPU, and a --cpu-offload option is provided for large-scale requests.
  • Supported Tools and Libraries:
  • ComfyUI: Official tutorials and workflow templates for R2V, FL2VA, and T2V are provided. You can load the included BF16 LoRA weights and execute them on nodes. – Hugging Face Diffusers: Official Diffusers pipeline documentation and sample code within the repository (examples/) are provided, allowing generation by calling directly from scripts.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — minimax-h3-community-license-agreement (Commercial use prohibited): MiniMax’s MiniMax H3 Community License. Use of the weights is limited to the EU, the UK, South Korea and the US (Sections II and V.4); elsewhere, including Japan, a separate license is required.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

This model’s license (minimax-h3-community-license-agreement) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-03 FastVideo NVFP4 FastVideo/Pika-H3-Flex-NVFP4-FP4Attn-QAD –
2026-11-03 FastVideo NVFP4 FastVideo/Pika-H3-Flex-NVFP4-QAD –

In addition, 2 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources