ABAB News Logo
Back to Flash News

Pika releases H3 Flex: open-weight video model reaches No. 1 on DesignArena’s text-to-video open-weight leaderboard

Pika launched Pika H3 Flex on October 29, 2026, a video model post-trained from MiniMax H3. The company said the model became No. 1 among open-weight models on DesignArena’s text-to-video leaderboard, ranking No. 7 overall for video quality, ahead of base H3, H3Max, and Seedance 2.5.

The model supports text-to-video, image-to-video with first and last frame images, multimodal reference-to-video, and reference-guided editing, and can take up to 12 reference inputs (images, video, audio). It generates 15-second 768p clips with synchronized stereo audio, and the company said the transformer computation time is as short as about 7 seconds; local open-source inference supports native 480p and 768p, while Pika’s hosted API provides 1080p, 2K, and 4K upsampling. The API can also export up to 5 separate audio tracks (dialogue, music, sound effects, foley, ambient sound). The company said the model is on average 1.8 times faster than H3Max and up to 80% cheaper.

Pika is simultaneously open-sourcing the model weights and inference code, which are available on Hugging Face and GitHub, and it supports ComfyUI and Diffusers. Local inference performance depends on hardware and implementation, and is not the same as hosted API speed. The model is accessible through Pika API Club, and is planned to enter Pika’s creation product on November 2. The company said post-training compressed the sampling steps from 50 steps to 6 steps, and that it did so through reinforcement learning and preference optimization.

Original sources:

  1. Title: Introducing Pika H3 Flex Publisher: Pika Publication date: 2026-10-29 URL: https://pika.art/blog/pika-h3-flex Nature of source: official announcement Verification status: full text read in this round, verification date 2026-10-29 Supported content: all facts in the body, including the model being post-trained from MiniMax H3, the No. 1 open-weight ranking on DesignArena and overall No. 7 ranking, about 7 seconds to generate a 15-second 768p clip, up to 12 references, 5 audio tracks, open weights and code, API and product availability, and so on

  2. Title: Pika H3 Flex — Faster-than-real-time video model Publisher: Pika Publication date: 2026-10-29 URL: https://pika.art/pika-h3-flex Nature of source: official product page Verification status: full text read in this round, verification date 2026-10-29 Supported content: 1.8x speed and up to 80% price advantage, resolution specifications, reference limits, audio track details, and open-source description (consistent with the blog)

  3. Title: Pika-H3-Flex/README.md Publisher: Pika-Labs (GitHub) Publication date: 2026-10-29 URL: https://github.com/Pika-Labs/Pika-H3-Flex Nature of source: official code repository Verification status: read in this round, verification date 2026-10-29 Supported content: open weights and inference implementation, local 480p/768p support, six-step generation, and adapter description

ABAB AI Insight

Pika positioned H3 Flex as a model for professional iteration rather than one-off final outputs. The company’s co-founder and chief scientist Chenlin Meng said that most video models assume users will finish after the first generation, but professional creators iterate repeatedly; Flex supports that workflow through faster generation, multiple reference inputs, and separate audio tracks, allowing selective upsampling to 4K after a 768p draft. Post-training compressed MiniMax H3’s 50-step sampling to 6 steps, then applied reinforcement learning and preference optimization, and the company said it surpassed both base H3 and H3Max on DesignArena.

The open weights cover the generation adapters and inference code for text-to-video, first-and-last-frame, and multimodal reference tasks; locally, it can generate 480p and 768p videos with synchronized audio. 4K upsampling, audio-track separation, and acceleration services remain on Pika’s hosted API. On pricing, the company said API prices start as low as $0.01 per output second, and that it has a significant cost advantage over H3Max. This release is consistent with Pika’s recent product moves: over the past three months, it launched function-first creation products, a lower-markup API platform, and a ChatGPT plugin.

Compared with closed models, open weights let developers run them on their own hardware, customize them, and build tools, but local speed and quality depend on hardware and optimization and are not the same as the controlled benchmark of the hosted API. The DesignArena ranking is a snapshot at release, and the company explicitly noted that the 7 seconds refer to transformer computation time rather than end-to-end API latency. The model is already available through API access and is planned to enter the creation product.

The indicators to watch are whether the subsequent DesignArena ranking holds and whether the open-source community actually adopts the six-step adapter and reports local performance.

ABAB News · Cognitive Law

  1. Video models are shifting from one-off generation to supporting repeated adjustment, with draft speed and separable audio becoming conditions for iteration.

  2. Open weights cover the generation adapters, while high resolution and audio-track separation are still provided by hosted services.

  3. Post-training compresses 50 steps to 6 steps, quality ranks above the base model, and both speed and cost decline.