Pika released Pika H3 Flex this week, a post-trained version of MiniMax H3 built for fast iteration. Pika’s approach is to explore ideas at low resolution, then finish the chosen take at higher resolution. The model is available in the Pika app, through the Pika API, and as open weights on Hugging Face and GitHub.
What changes
- Open weights, with limits. The release includes inference code for text-to-video, first/last-frame video and reference-to-video, with audio, at 480p and 768p. A ComfyUI workflow is provided. The 1080p, 2K and 4K upsampler, audio-stem separation and accelerated serving stay on Pika’s hosted service. The product page does not name a license.
- More references. One job can use up to 12 references: up to nine images, three videos and three audio clips. Prompts can run to 8,000 characters.
- Clips and audio. Output runs 3 to 15 seconds at 24 fps in 11 aspect ratios, with synchronized stereo audio. Dialogue, music and effects can be returned as separate WAV stems.
- Speed. Pika says a 15-second 768p clip can take as little as seven seconds of model compute, and that Flex is about 1.8 times faster than H3Max on average. That figure is not end-to-end API latency.
Prices
On the hosted API, output costs $0.02 per second at 480p, $0.04 at 768p, $0.07 at 1080p, $0.08 at 2K and $0.16 at 4K. Launch prices are half of that. On Pika’s Create app, one credit equals $0.01, so a 15-second 768p clip uses 60 credits at the regular rate. Pika API Club members and Creator and Fancy subscribers get 15 free generations a day until November 15.
What it means for users
Video teams that make many short variations can draft cheaply at 480p and upscale only the takes they keep. Developers who want to run the model themselves can test the open weights locally or in ComfyUI. Plan for hosted calls if you need 4K output or separate audio stems. Pika also says local speed depends on hardware and setup and may not match its API.
Pika reports that Flex ranks seventh for video quality on DesignArena’s text-to-video board, with a score of 1,293. That is the company’s reading of a public leaderboard, not an independent test.
Related on ToolAI
See the Pika profile, and compare it with Runway and Luma AI. To run the open weights in a node workflow, start with ComfyUI.
Source: Pika: H3 Flex