Baseten

Baseten is an inference platform for running open-weight LLMs and custom models in production, with per-token Model APIs, dedicated GPUs and training.

Baseten gives engineering teams a hosted place to serve language, speech and embedding models behind an API, so they never have to run a GPU fleet themselves. It belongs to Baseten Labs, Inc., an independent San Francisco company that started the platform in 2019. Underneath sits the Baseten Inference Stack, the company's own model runtimes, with autoscaling, fast cold starts and availability across several clouds. Serving, scaling and training share one account. Model APIs offer ready endpoints for well-known open-weight models, including the GLM, DeepSeek and Kimi families. Every Model API supports tool calling, and structured outputs return JSON that matches a supplied schema. Vision models accept images in the request, and some models expose extended thinking. Dedicated Deployments put any model on reserved GPUs, with asynchronous requests and priority queuing for slower jobs. Baseten Embeddings Inference covers embedding, classification and reranking, while the speech stack handles transcription, speaker diarization and audio understanding. The platform suits product teams adding AI features for real users, startups taking a fine-tuned model to production, and enterprises bound by data rules. Synchronous requests are not stored by default, and regional environments keep replicas and traffic inside a chosen geography. Usage is metered per token on Model APIs and per minute of GPU time elsewhere, starting on a pay-as-you-go plan. Pro, sold on a quote, adds priority GPU access and dedicated compute. Enterprise adds custom SLAs, self-hosting in a customer's VPC or a hybrid setup, and RBAC with Teams. Typical work includes hosting a chatbot on an open model, transcribing recorded calls, or running reranking behind a search feature. Because the endpoints are compatible with the OpenAI SDKs and the Anthropic API, application code written for either can call them with little rework. Training and Loops fine-tune a base model on a company's own data, on the same GPU types offered for deployments. Baseten Switch connects Claude Code, Codex CLI or Pi to models hosted on the platform. That makes Baseten a fit for teams that want open models in production under their own control.