purplesquirrelmedia.io

What It Is

Purple Squirrel Media’s PSM-V3 is a multimodal foundation model for video generation, aimed at production use and offered in three ways: API access, on-prem private deployment, and open weights. The page claims support for text, image, audio, and video inputs, producing cinematic-quality video with specs including native 4K, up to 60fps, and generations up to 120 seconds long.

Core Capabilities

In terms of coverage, PSM-V3 goes beyond text-to-video and also includes image-to-video, audio-driven video, and video-to-video re-rendering. On the audio side, it can drive visual pacing, transitions, and scenes based on music rhythm, speech, and ambient sound. Video-to-video supports style transfer, super-resolution, and frame interpolation. The architecture details mention a multimodal encoder, a 3.2B-parameter Diffusion Transformer, Temporal Attention, a VAE Decoder, and C2PA signing. On the ecosystem side, it supports REST API access, streaming output, webhooks, and batch processing, with integrations listed for ComfyUI, Hugging Face, Replicate, RunPod, NVIDIA NIM, fal.ai, Together AI, and others.

Pricing and Deployment

The page clearly states that there is a free tier with no credit card required, but it does not disclose the free quota. The hosted API is billed per generation and supports volume discounts. Enterprise plans are settled via purchase orders and Credits, with 1 Credit equal to USD 1 in value. The open weights are released under Apache 2.0, with fine-tuning scripts and LoRA support provided. Self-hosting has a high hardware bar: the minimum requirement is an NVIDIA A100 80GB or H100, while dual H100s are recommended for real-time 4K generation.

Pros, Cons, and Limitations

The strengths are its broad input modality support, flexible deployment options, reduced lock-in risk through open weights, and support for private VPC/data center deployment to meet data sovereignty requirements. The downsides are that the page does not provide specific unit pricing, a real sample gallery, third-party benchmarks, or failure cases. Claims such as “production-ready” and “low latency” still need to be validated in real-world testing. Details around hosted API data retention and whether data is used for training are also not fully disclosed in the main content.

Who It’s For and Access from China

It is better suited to video generation product teams, advertising and film/TV concept teams, music visualization apps, and enterprises with GPU resources that need fine-tuning. It is less suitable for individual users who only want to generate short videos occasionally at low cost. Access from mainland China, supported payment methods, and Chinese-language support are not disclosed, so real-world use may require testing network connectivity. Comparable alternatives include Runway, Pika, Luma, Kling, PixVerse, Sora, and others.