Reflection AI's Beam Reasoning Model Goes Live on AI Gateway, Free During Launch

A model that once lurked in testing under the codename stealth/cat-jam has officially surfaced. Reflection AI's reasoning model Beam is now available on Vercel's AI Gateway and is free for all developers during the launch period. This isn't your typical "limited-time trial" gimmick—the weights themselves are open-sourced under Apache 2.0, so anyone can call it freely or even self-host it.

What Is Beam: An Open-Weight Reasoning Model Built for Speed

Beam's positioning is clear: a low-latency American open-weight reasoning model, primarily serving coding and agent tasks. It supports tool calling and structured outputs, and its context window reaches 1 million tokens—enough to swallow an entire code repository, long documents, or a complete conversation history.

What truly sets it apart from other open-source models at its level is its architectural trade-offs. Beam has 501B total parameters, but only activates 23B per token—a classic MoE (Mixture of Experts) approach. More importantly, it "actively skips tokens the task doesn't need," performing dynamic pruning along the reasoning chain to deliver answers faster than larger open models. Reasoning effort is adjustable, meaning developers can balance between "fast and shallow" and "slow and deep" to fit scenarios ranging from code completion to multi-step agent planning.

MoE (Mixture of Experts) is a sparse-activation model architecture: the model internally consists of multiple "expert" subnetworks, and during each forward pass, a routing mechanism selects only the few experts that best match the current token to participate in computation, leaving the rest dormant. This is exactly why Beam can have "501B total parameters while activating only 23B per token." Unlike dense models that must use all parameters every time, MoE decouples parameter count from per-inference computational cost: parameter count determines the model's knowledge capacity and expressive power, while activation count determines actual compute and latency overhead. The trade-off is that VRAM usage is still calculated based on total parameters, since all experts must reside in memory; routing itself also introduces some additional overhead, and imbalanced loads can lead to some experts being undertrained. For developers, understanding this helps gauge deployment barriers—Beam's VRAM requirements are close to a 501B-class model, while its compute requirements are close to a 23B-class model.

Two Free Access Paths with Different Expiration Behavior

Reflection AI offers two free paths, differing in what happens after the free period ends:

Model Name	After Free Period Ends
reflection/beam	Automatically transitions to standard billing, zero downtime
reflection/beam-free	Service stops

If you're evaluating this model for production, opting for the standard name reflection/beam is safer—once the free period ends, it smoothly switches to paid without suddenly cutting off. For developers who just want short-term experimentation and don't care about long-term availability, reflection/beam-free is more hassle-free—use it and it's done.

Invocation is straightforward. Here's an example using the AI SDK:

import { streamText } from 'ai';

const result = streamText({
  model: 'reflection/beam',
  prompt: 'Outline three steps for debugging an API timeout.',
});

for await (const text of result.textStream) {
  process.stdout.write(text);
}

Mainstream coding agents like Claude Code, Codex, and Cursor can also connect directly—just install the latest Vercel CLI, run npx vercel ai-gateway setup, then select the corresponding model name in your agent.

AI Gateway is a proxy service that sits between applications and model providers. Applications send requests only to the Gateway's unified interface, and the Gateway routes requests to specific model providers while handling authentication, retries, failover, rate limiting, and usage metering in between. This design brings several practical benefits: first, interface normalization—when switching or adding models, you only need to change the model name, not rewrite SDKs and auth logic; second, enhanced reliability—when a provider times out or errors out, the Gateway can automatically retry or reroute to a backup channel, and its availability SLA is often higher than that of any single provider; third, observability—call volume, token consumption, and costs can be centrally tracked and reported. Zero Data Retention means request and response content is not persistently stored and is used only for the current inference, which is especially critical for enterprises dealing with sensitive code or compliance requirements.

The Value of AI Gateway: Unified API and No Markup

Beam choosing to launch on AI Gateway rather than standalone hosting is itself a smart move. AI Gateway provides a unified API to call models, track usage and costs, and has built-in retries, failover, and performance optimization—with the goal of achieving higher availability than the model providers themselves. It also supports custom reports, Zero Data Retention, API key budget controls, and routing rules.

For cost-sensitive teams, AI Gateway's pricing strategy is worth noting: pass-through at provider list price, no platform fee, no markup on inference costs, and even BYOK (Bring Your Own Key) requests incur no extra charges. This means calling Beam through it won't be more expensive than connecting directly—what you get extra is a layer of reliability assurance and observability.

Open Weights + Free Access: What Signal Does This Send?

By releasing Apache 2.0 weights and a 1M-context reasoning model directly, Reflection AI's intent goes beyond "letting more people use it." Open weights mean enterprises can deploy privately, fine-tune, and audit—a hard threshold for data-sensitive teams. And Beam's efficiency-focused approach of "activating 23B parameters to go fast" is effectively drawing a line against giant models that require massive VRAM and expensive inference: it's not about bigger parameters being better, but about getting answers right within a given latency budget.

For developers, the lowest-cost way to validate right now is to plug it into your existing agent workflow during the free period. Connect using the standard name reflection/beam, evaluate, then decide whether to switch to paid—migration cost is nearly zero.

Apache 2.0 is one of the most permissive open-source licenses in the industry. It allows users to freely use, modify, and distribute the software, including for commercial closed-source products, and does not require derivative works to be contributed back under the same license (unlike GPL's copyleft provisions). It also includes explicit patent grant clauses, reducing the risk of patent litigation for users. For "open-weight" models, adopting Apache 2.0 means enterprises can deploy them privately on internal networks, fine-tune and distill based on their own data, and conduct security audits of model behavior—without worrying about license contagion or commercial restrictions. This is also why license choice is often seen as one of the key thresholds determining whether an open-weight model can truly be adopted by enterprises—by contrast, some custom licenses with usage or commercial restrictions, even if weights are downloadable, still face legal review during deployment.

One-line Summary

Reflection AI's open-weight reasoning model Beam lands on AI Gateway, free during launch.

Paragraph Summary

Reflection AI's reasoning model Beam is now available on Vercel's AI Gateway and is free during the launch period, with weights open-sourced under Apache 2.0. Beam is positioned as a low-latency American open-weight reasoning model for coding and agent tasks, supporting tool calling, structured outputs, and adjustable reasoning effort, with a 1 million token context window. Its architecture uses an MoE approach: 501B total parameters but only 23B activated per token, with dynamic skipping of unneeded tokens. Free access offers two paths: reflection/beam and reflection/beam-free. The former auto-transitions to paid with zero downtime, while the latter stops serving upon expiration. With AI Gateway's unified API, pass-through pricing without markup, and Zero Data Retention, developer migration costs are minimal, allowing quick integration with mainstream coding agents like Claude Code and Cursor for validation.