FriendliAI

FriendliAI provides inference infrastructure for developers and companies running open-weight and custom generative AI models. Its Model API gives applications access to models for coding and agent workflows, while Dedicated Endpoints deploy customers’ own models or models from Hugging Face on dedicated GPU instances without infrastructure management. Bring Your Own GPU (BYOG) puts the inference stack in customers’ cloud environments, where FriendliAI handles deployment, monitoring and scaling while data stays within that cloud.

Founder and CEO Byung-Gon Chun established FriendliAI with members of his Seoul National University research team. Chun led the team behind continuous batching, introduced in the 2022 Orca paper: requests are scheduled at each inference iteration to improve throughput and GPU utilization as workloads change. FriendliAI combines this scheduling approach with custom kernels, speculative decoding and reuse of previously computed context through KV cache optimization. These techniques address the computation and latency costs of serving models through repeated agent steps.

The company identifies Twelve Labs and LG as production customers and expanded its U.S. presence with a San Francisco office in 2026. Its FriendliLink command-line tool connects existing coding agents, including Claude Code, Codex and Cursor, to Friendli Model APIs. Developers can discover models and prices, switch models and restore their previous configuration through one interface.