Reflection AI, a startup backed by Nvidia and founded by two former Google DeepMind researchers, introduced Beam in October 2026: an open-weight model with 501 billion total parameters and only 23 billion active per token. It is built for coding and agentic work with a one-million-token context window. To help developers test it, Vercel is offering free launch-period access to Beam through its AI Gateway.

Before you begin, you will need a free Vercel account and the latest version of the Vercel CLI installed on your machine. The free offer applies only when Beam is served by Reflection AI through the gateway, so create an API key on the Vercel AI Gateway page and keep it handy before running any setup commands.

Install or update the CLI, then run the guided setup command. Open a terminal and type the setup instruction Vercel documents for coding agents. When prompted, select reflection/beam or reflection/beam-free as your model. The plain name automatically transfers to paid pricing when the free period ends, while the free-suffixed name stops serving instead — choose the one that matches your intent.

Once selected, you can point your coding agent at Beam inside tools like Claude Code, Codex, or Cursor. Beam supports tool calling and structured outputs, so it slots into existing agent loops that call functions or edit files. Because it activates only a fraction of its parameters per token, it tends to reach answers quickly on coding and agentic tasks.

You can also call Beam through the Vercel AI SDK using its generateText and streamText functions, or through the OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages APIs by changing the base URL. Set your gateway key as AI_GATEWAY_API_KEY in your environment. Vercel adds no markup on inference pricing, so hosted costs pass through directly.

Keep two caveats in mind. First, the free period is temporary, so treat it as an evaluation window: run Beam against your own documents and codebases before standardizing on it. Second, Reflection has also released the Beam weights under Apache 2.0 for self-hosting, but served weights and self-hosting are different arrangements — open models still need proper evaluation before they earn a place in your stack.