Pika Releases Pika H3 Flex Video Model With Open Weights
Post-trained on MiniMax H3, the new model condenses generation into six steps, introduces multi-reference directing, and provides open weights for local pipelines.
In short
- Pika released Pika H3 Flex on October 29, 2026, post-training MiniMax H3 down from 50 sampling steps to six.
- The open-weight release provides local 480p and 768p generation with synchronized 32 kHz stereo audio across 11 aspect ratios.
- A single generation can ingest up to 12 multimodal references, spanning images, video clips, and audio tracks.
- Pika offers hosted API options starting at $0.01 per second for text-to-video, supporting 4K finishing and five isolated audio stems.
- In Pika's evaluations, the model achieved the top open-weight ranking on DesignArena's text-to-video leaderboard.
Pika Labs launched Pika H3 Flex on October 29, 2026, delivering an open-weight video generation model designed to accelerate iterative creative production. Built as a post-trained derivative of MiniMax H3, the architecture compresses generation to six sampling steps while producing synchronized audio alongside motion. Pika released the model adapters and inference code publicly, allowing developers and studios to run generation pipelines locally on their own compute or access extended resolution and audio tools through hosted endpoints.
According to the Pika H3 Flex Technical Deep Dive, the system combines fine-tuning with reward-based optimization on an aesthetic dataset curated internally. Pika structured the release around the iterative loop of production design, prioritizing quick revision and prompt responsiveness over single-pass, slow-rendering workflows.
Six-Step Generation and Multimodal References
The central architectural adjustment in Pika H3 Flex is the reduction of diffusion steps. Standard MiniMax H3 workflows typically require 50 sampling steps to resolve a scene. Pika condensed this process to six generator evaluations, allowing a 15-second clip at 768p resolution to complete in as little as seven seconds of transformer computation, according to benchmarks reported by Superpower Daily. Pika clarified that this seven-second measurement reflects Diffusion Transformer (DiT) computation rather than end-to-end request turnaround, which includes conditioning overhead and decoding.
To control output composition, the model splits conditioning across two specific task adapters: FL2VA and Ref2VA. As detailed on the Pika H3 Flex Hugging Face repository, FL2VA governs text-only generation and endpoint keyframing, including first-frame, last-frame, and combined first-and-last-frame conditioning. Ref2VA handles multimodal reference conditioning, accepting up to 12 reference files in a single pass.
These 12 inputs can mix up to nine images, three video clips, and three standalone audio files. Each video or audio reference must run between 2 and 15 seconds, with total reference video and total reference audio capped at 15 seconds each. This conditioning allows users to assign distinct roles to different assets within a single text prompt—such as defining character appearance with one image, lighting with another, and pacing from an uploaded motion sequence.
Technical Specifications and Local Pipelines
Pika H3 Flex operates locally across 11 aspect ratios, including 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16, and 9:21. The open model supports native rendering at 480p and 768p resolutions at 24 frames per second. Output durations range from 5 to 15 seconds, mapped to an underlying frame grid where a 5-second request yields 124 frames and a 15-second request renders 362 frames.
Local generations produce an MP4 file containing H.264 video and 32 kHz stereo AAC audio. As noted by ArtRealmAI, both FL2VA and Ref2VA are distributed as rank-128 FP32 LoRA adapters. Community formats, including BF16 LoRAs for ComfyUI and integrations for Hugging Face Diffusers, were published simultaneously alongside standard command-line tools.
Running the model locally requires substantial enterprise compute. The underlying MiniMax H3 base components consume roughly 210 GB of storage. Pika verified CLI execution on Nvidia H100 80GB GPUs, noting that complex reference requests and larger canvas sizes necessitate host-memory offloading via command-line flags. The adapter weights and base components remain governed by the MiniMax H3 Community License Agreement, while the repository's example execution code and Qwen text-encoder components carry Apache 2.0 licensing.
API Features, Audio Stems, and Pricing
While local weights support draft iterations up to 768p, Pika reserved finishing capabilities for its hosted infrastructure. As outlined on VP Land, the hosted Pika API includes a draft-to-video endpoint that upscales draft renders into 1080p, 2K, and 4K outputs.
The hosted API also introduces audio demixing that exports five distinct stems: dialogue, score, sound effects, foley, and ambience. This separation allows editors to adjust speech clarity, mute background music, or rebalance environmental effects without leaving the video generation environment.
Commercial API access is available through Pika API Club, with text-to-video pricing starting at $0.01 per second of generated footage. Reference-to-video generation is priced at $0.03 per second. During an initial two-week launch window, Pika is providing 15 free generations per day. The capabilities are scheduled to reach Pika's consumer web creation application on November 2, 2026, as reported by Frame Fandom.
Benchmark Performance and Industry Context
Pika reported that H3 Flex took first place among open-weight models and seventh place overall on DesignArena's text-to-video leaderboard, recording a score of 1293. This places it ahead of base MiniMax H3, fal's H3 Max, and Seedance 2.5 on that specific benchmark. In Pika's internal human evaluation study, Pika claims H3 Flex ranked highest in human anatomy representation and artistic style fidelity.
The release adds to a growing ecosystem of specialized models built on top of MiniMax's foundation. Third-party variants such as fal's H3 Max and OpenVDN's VDN-H3 have similarly explored accelerated sampling and architectural modifications. By providing open weights, prompt-driven shot direction, and modular audio outputs, Pika H3 Flex establishes a dual track that bridges local studio pipelines with hosted high-resolution finishing.
Questions people ask
What is Pika H3 Flex?
Pika H3 Flex is an open-weight omni video generation model developed by Pika Labs. It is post-trained on MiniMax H3 to generate 5- to 15-second video clips with synchronized stereo audio in six sampling steps.
Can Pika H3 Flex be run locally?
Yes. Pika released open weights and inference code for Hugging Face Diffusers, ComfyUI, and a CLI tool. Local inference supports native resolutions of 480p and 768p, though the base model assets require approximately 210 GB of disk space and enterprise hardware like an Nvidia H100 GPU.
How does the model handle audio and higher resolutions?
Local inference generates mixed 32 kHz stereo audio within the MP4 file. Higher resolutions—including 1080p, 2K, and 4K—as well as five exportable audio stems (dialogue, score, SFX, foley, and ambience) are available through Pika's hosted API.
How much does Pika H3 Flex cost via the API?
Hosted API pricing begins at $0.01 per output second for standard text-to-video generation and $0.03 per second for reference-conditioned video generation.
Sources
- Efficient-Large-Model/Open-H3-Regenerate-2KHugging Face, Oct 26, 2026, official source
- Pika H3 Flex Technical Deep DivePika, Oct 29, 2026, official source
- Pika-Labs/Pika-H3-FlexHugging Face, Oct 29, 2026, official source
- Pika releases open-weight video model with six-step generation and synchronized audiosuperpowerdaily.com, Oct 30, 2026
- Pika H3 Flex Is a New Open-Weight Video Model Built for Fast Iteration, With Output Up to 4KVP Land, Oct 29, 2026
- Pika H3 Flex: a fast open-weight video model with 12 references and split audio stemsWren, Oct 29, 2026
- Pika H3 Flex: Open Six-Step Video and Audio ModelArtRealmAI, Oct 29, 2026
How this story was made: the newsroom picked it up from Hugging Face, Google News and Reddit, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (20 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 30, 2026 at 03:58 UTC
