Pruna AI promises five-second video clips in two seconds

Why it matters

Open-weight video models let infrastructure startups compete through inference speed and pricing. Pruna AI is betting developers will pay for an optimized H3 endpoint instead of operating the model themselves.

Pruna AI launched P-Video-2-Pro in a thread on X on September 17th, giving developers a faster commercial endpoint for generating short videos with audio from text or reference images.

https://x.com/PrunaAI/status/2100585945988833736

Rayan Nait Mazi, Bertrand Charpentier, John Rachwan and Stephan Gunnemann founded Pruna AI around model efficiency rather than training foundation models from scratch. Nait Mazi serves as CEO, Charpentier as president and chief scientist, Rachwan as CTO and Gunnemann as chief strategy officer. The founders built the research-heavy operation between Munich and Paris after work connected to the Technical University of Munich.

P-Video-2-Pro is based on MiniMax H3, the multimodal video model that MiniMax released in late July and opened on August 3rd. Pruna AI's move less than seven weeks later illustrates the commercial opening created by open-weight models: independent infrastructure teams can optimize the serving stack, narrow the interface and sell speed without paying to train the underlying foundation model.

What Pruna AI shipped

According to the P-Video-2-Pro documentation, the endpoint accepts a text prompt, a first-frame image or a combination of first- and last-frame images. It returns clips lasting between five and 15 seconds at 24 frames per second, with generated audio included.

Developers can choose 480p or 768p output and seven aspect ratios, including 16:9, 9:16 and square video. A seed parameter allows repeatable runs. Pruna AI also exposes three prompt-expansion settings: off, turbo and max.

The endpoint has separate speed and quality modes. Pruna AI says speed mode can generate a five-second 480p clip in about two seconds and a five-second 768p clip in roughly 4.3 seconds. Those are Pruna AI's latency measurements. The launch materials do not provide independent testing that would substantiate Pruna AI's broader claims of offering the "best quality" and "best speed."

P-Video-2-Pro's interface exposes a smaller slice of H3 than MiniMax's full release. MiniMax says H3 can generate resolutions up to 2K and accept combinations of images, video and audio as references. Pruna AI caps this endpoint at 768p and does not expose audio input. That narrower product is built around the tasks developers are likely to put into automated pipelines: text-to-video, image-to-video and first-to-last-frame transitions.

Speed is the business model

Pruna AI lists the 480p speed mode at $0.02 per second of finished video and 480p quality at $0.04. At 768p, the rates rise to $0.035 for speed and $0.075 for quality. A five-second clip therefore carries a list price ranging from $0.10 to $0.375, depending on resolution and mode.

Pruna AI is discounting those rates by 50% through September 24th, 2026. The API permits as many as 250 requests per minute, positioning P-Video-2-Pro for application developers rather than creators generating clips individually through a consumer interface.

The distinction matters because AI video economics depend on iteration. A cheap preview is useful only when developers can generate enough versions to test prompts, camera directions and scene continuity before paying for the final output. Pruna AI's two-mode setup turns that workflow into the pricing model: speed mode handles experimentation, while quality mode charges twice as much or more for production runs.

Pruna AI has spent its first years selling the machinery behind that tradeoff. Its open-source optimization framework combines techniques including caching, quantization, pruning, distillation and hardware-specific compilation. The framework supports language, image, video and speech models, and its GitHub repository has accumulated about 1,300 stars.

The founders raised a $6.5 million seed round in November 2024 led by EQT Ventures, with Daphni, Motier Ventures and Kima Ventures participating. At the time, Pruna AI's pitch centered on making third-party models faster, smaller and cheaper. P-Video-2-Pro carries that thesis into a product developers can buy by the second.

That shift gives Pruna AI control over packaging, pricing and distribution rather than leaving its optimization work buried inside another provider's infrastructure. It also creates a harder test. Developers can compare the endpoint's latency and output directly against MiniMax H3 services and the growing list of video APIs built on open models.

Pruna AI's launch numbers make speed the immediate selling point. Sustained usage will depend on whether that acceleration preserves motion, prompt adherence, audio synchronization and character consistency across repeated generations. In video inference, saving seconds matters only when the output avoids another paid rerun.