FriendliAI, founded in 2021 and spun out of Seoul National University, develops infrastructure to accelerate the deployment of large language models. The company commercializes research in AI inference optimization, focusing on techniques such as custom GPU kernels, caching, continuous batching, speculative decoding, and parallel inference. Its primary platform claims to deliver over twice the inference speed of standard solutions and provides access to more than 590,000 models from the Hugging Face repository for deployment.
The company's product suite includes a general AI inference platform, Dedicated Endpoints for enterprise clients running custom or fine-tuned models, and container-based solutions. Both the Dedicated Endpoints and container solutions are offered with a 99.99% uptime service level agreement. FriendliAI operates from Seoul and San Francisco.
