fal documented how it post-trained MiniMax H3 to produce H3 Max, then deployed the video model on fal Serverless and exposed it through a Model API. Its deployment report separated time spent waiting for a GPU worker from time spent generating the video, showing where serving and model optimizations affect latency.
fal.ai
Also known as: fal · fal Serverless · fal Model APIs
Plain English
fal.ai runs AI models for generating and editing media. Developers can call a hosted model or deploy their own model or workflow without managing the GPUs themselves.
Technical Overview
fal’s Model APIs and custom Serverless deployments share a managed inference runtime and request queue. Clients can submit long-running jobs and retrieve results by polling or receiving a webhook. The queue can absorb bursts while the platform starts GPU workers, so total latency includes time spent waiting as well as time spent generating.
Chronological Entries (4)
fal introduced a Model Context Protocol server for its model catalog. It exposes separate tools for submitting a long-running job and checking its status, so a client can start a video generation request and retrieve the result later.
fal introduced a hosted ComfyUI runtime. Developers could run node-based workflows on managed GPUs and expose those workflows as APIs, allowing an application to invoke the pipeline without opening the editor.
fal introduced managed web endpoints for Python functions running on fal-serverless. Developers could expose a deployed function as a REST API while fal managed the web server that routed requests to it.