FriendliAI

FriendliAI runs open-weight and custom LLMs for teams building AI agents, through an OpenAI-compatible inference API billed by usage or by the GPU hour.

FriendliAI takes the GPU serving work off a team that wants a large language model inside its product, from the first API call to a fleet of dedicated GPUs. FriendliAI Corp, an independent company based in San Francisco with a hub in Seoul, Korea, builds and runs it. The platform hosts models from families such as GLM, DeepSeek, MiniMax and Gemma rather than training its own. Every part runs from one account in the Friendli Suite console. Model APIs give pay-per-token access to a catalog of hosted models for text, vision and speech to text, with Whisper handling audio transcription. Requests support tool calling, structured outputs that follow a supplied JSON schema, and reasoning controls set by effort or by token budget. Dedicated Endpoints put a Hugging Face or custom model on single-tenant GPUs. They add autoscaling, versioning, online quantization and multi-LoRA serving of several adapters on one base model, and they also serve embeddings. BYOG runs the same Friendli inference stack on GPUs the customer already owns. The fit is developers and AI teams shipping agents, coding assistants and chat features on open-weight models. Teams with their own trained models or LoRA adapters use Dedicated Endpoints to serve them without building a cluster. Companies that own GPU fleets take BYOG, while enterprises needing a VPC, an on-premises install or a chosen region sign the Enterprise plan. Model APIs are paid per token from prepaid credits, and Dedicated Endpoints by the GPU hour, metered by the second with nothing charged for idle time. BYOG and Enterprise are priced by contract. Because the API follows the OpenAI format, an application written against the OpenAI SDK in Python or Node.js calls FriendliAI through the client library it already uses. An agent built with LangChain sends its tool calls to a FriendliAI model, and Weaviate connects to it for retrieval work. A model trained elsewhere goes from its Hugging Face repository to a running endpoint in the console. The jobs it is hired for are serving a coding or support agent, transcribing recorded audio, and running a fine-tuned model in production without operating the GPU cluster underneath it.