Reflection AI's Apache 2.0 Beam reasoning model is free on Vercel's AI Gateway

Beam lands on Vercel's AI Gateway

Beam, an open-weight reasoning model from Reflection AI aimed at coding and agentic tasks, is now available through Vercel's AI Gateway, and according to Vercel it is free to use for a limited launch period. Before this public release, the model had been reachable on the platform in stealth under the name stealth/cat-jam.

Sparse by design

Vercel describes Beam as a low-latency, US-built, open-weight reasoning model whose weights are published under the Apache 2.0 license. It supports tool calling and structured outputs, offers a 1M-token context window suited to code, documents, and long conversation history, and lets developers tune how much reasoning effort it applies.

The distinctive engineering choice is sparsity. Beam is a 501-billion-parameter model that activates only 23 billion parameters per token, skipping computation that a given task does not need. Vercel says the result is that Beam often arrives at its answer faster than larger open models do.

Two routes during the free period

Vercel offers two ways to use Beam at no cost, each behaving differently once the launch offer ends:

• reflection/beam is the standard model name. When the free window closes, this slug automatically moves over to paid pricing with no downtime, so existing integrations keep working.
• reflection/beam-free carries a -free suffix. This variant simply stops serving when the offer lapses.

To use the model inside agent tools such as Claude Code, Codex, or Cursor, Vercel says to install the latest Vercel CLI and run npx vercel ai-gateway setup, then select either slug from within the agent. Beam can also be tried directly in Vercel's model playground.

What the gateway adds

AI Gateway is Vercel's unified API for invoking models. Beyond routing requests, it handles usage and cost tracking, retries, failover, and performance optimizations that Vercel positions as delivering better uptime than connecting to a provider directly. It ships with custom reporting, Zero Data Retention support, budgets scoped to individual API keys, and configurable routing rules. Vercel states that gateway pricing mirrors provider pricing without markup, and that no platform fee is charged on inference, including for bring-your-own-key requests.

Why it matters

Beam's architecture targets the bottleneck that matters most for coding agents: latency accumulated across many small tool-calling steps. A very large parameter pool combined with aggressive per-token sparsity is a bet that interactive agents gain more from arriving at answers quickly than from maximizing activated capacity on every single token. The 1M-token window and adjustable reasoning effort are similarly aimed at agentic workflows that juggle large codebases and long-running conversations.

The Apache 2.0 release also gives teams a genuine choice: run and fine-tune the weights themselves, or take the managed route through a gateway that layers on failover, usage budgets, and no-markup inference. The two-slug launch scheme is a small but practical detail — evaluating under the standard name means moving to paid usage requires no code change and no downtime, while the -free variant makes the offer's expiry explicit in the identifier itself and stops serving rather than silently starting to bill.