Runware

Runware is an AI inference platform that puts image, video, audio, language and 3D models behind a single API, alongside serverless GPUs and reserved compute for teams that want to run workloads of their own. The company designs and operates its own modular data centers, called Sonic Pods, and pairs them with an orchestration layer that keeps models preloaded across regions so each request routes to hardware that already holds the weights in memory. The platform targets developers and product teams building generation features into applications. A request body is an array of tasks, so an image, a video and a chat completion can travel in one call, over REST for stateless work or WebSockets for persistent sessions, with a webhook per task when results should be pushed back. More than 400,000 models share one addressing scheme covering hosted, partner and community checkpoints as well as weights you upload yourself, and on open models you control every parameter down to stacked LoRAs and a swapped VAE. Model upload accepts checkpoints, LoRAs, LyCORIS, VAEs and textual inversion embeddings, with public or private visibility and version history. Pricing is pay as you go, billed per generation for media and per million tokens for language models. Published examples include Ideogram 4.5 at $0.008 per image, Qwen Image 2.1 Pro at $0.075 per image, video rates from $0.02 per second at 480p, and text models ranging from $0.10 to $10.00 per million input tokens. Serverless GPU compute is metered by the second at $1.99 per GPU hour and scales to zero when idle. New accounts registered with a business email receive $2 in free credits. Model upload is in beta and free of charge, after which storage will be billed at $0.05 per GB per month.