Reflection AI's Beam lands free on AI Gateway: 501B parameters, 23B active per token, built for low latency
Synopsis
Reflection AI's Beam reasoning model has left stealth testing under the codename stealth/cat-jam and is now available on AI Gateway under its own name, free for a limited launch period when served by Reflection AI. It targets coding and agentic tasks, ships weights under Apache 2.0, supports tool calling and structured outputs, and carries a 1M-token context window. The model has 501B total parameters but activates only 23B per token and skips tokens a task does not need, so it tends to answer sooner than larger open models.
Interpretation
A reasoning model previously tested on AI Gateway under the stealth codename stealth/cat-jam is now open under Reflection AI's Beam name, free during the launch period when serviced by Reflection AI. Before this, the model was reachable only under a stealth codename, with no public entry under a正式 name. Free access comes in two forms: the standard name reflection/beam transfers automatically to pricing with no downtime when the free period ends, while reflection/beam-free stops serving at that point.
Beam targets coding and agentic tasks, supports tool calling and structured outputs, and offers a 1M-token context window for code, documents and conversation history. Reasoning effort is adjustable, so users can choose how much reasoning to spend per task. The weights are released under Apache 2.0 as a US open-weight model.
Speed comes from sparse activation: of 501B total parameters, only 23B activate per token, and the model avoids tokens a task does not need. Larger open models typically involve more parameters in each token's computation. By Reflection AI's description, this design lets Beam tend to reach answers sooner than larger open models.
Perspective
The report covers the launch information Reflection AI published: Beam is free on AI Gateway under reflection/beam and reflection/beam-free, with the former moving automatically to pricing after the free period and the latter stopping service then. The model is positioned as a US open-weight reasoning model for coding and agentic tasks, with Apache 2.0 weights, tool calling, structured outputs, adjustable reasoning effort and a 1M-token context window. Access paths include the Vercel AI SDK streamText call and selecting the model in Claude Code, Codex, Cursor and similar agents after running the Vercel CLI ai-gateway setup command.
The speed claim is phrased as tending to reach answers sooner than larger open models, without naming comparison models, tasks or measurement. The length of the free period is not stated; only that the standard name moves to pricing and the -free name stops serving afterward. The specific mechanism of sparse activation and skipping unneeded tokens is not detailed in the material.