Reflection AI's Beam: a 501B open-weight coding model, free on Vercel AI Gateway

Reflection AI released Beam, its first open-weight model, under an Apache 2.0 license with the weights published on Hugging Face. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and only 23 billion active per token, aimed at coding, reasoning, and long-horizon agentic workloads. It had circulated in stealth as "stealth/cat-jam" before launch. The model also carries a 1M-token context window, and Vercel is serving it on AI Gateway for free during the launch period.

What Beam is, in numbers

Reflection's announcement describes the model as built for "frontier reasoning efficiency": the 501B-total, 23B-active design activates only a small expert slice per token, and the model skips tokens a task doesn't need, so it reaches answers sooner than larger open models. Beam supports tool calling, structured outputs, and adjustable reasoning effort (Vercel changelog).

Weights ship in BF16, FP8, and NVFP4 precision, plus a speculative decoding drafter model. Day-zero serving and training integrations include Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and NVIDIA Dynamo, so self-hosters can go from download to production endpoint without custom plumbing. A final benchmark set and a Safety Technical Report accompany the release.

The free window on Vercel AI Gateway

Vercel offers two access paths. The standard slug reflection/beam is free during the launch period and automatically transfers to paid pricing when the window ends, with no downtime. The reflection/beam-free variant also costs nothing now, but stops serving entirely when the offer ends. Both work through the AI SDK and Chat Completions, and the Vercel CLI can wire Beam into coding agents such as Claude Code, Codex, and Cursor.

Vercel says AI Gateway passes through provider pricing with no markup and charges no platform fee on inference, including Bring Your Own Key requests. Neither source publishes the exact end date of the free period, so treat it as open-ended and watch for a changelog update before you depend on the free tier.

The platform push behind the model

Reflection is launching with more than 40 partners and calls the effort an "open operating system for agents": the model is the kernel, and the layers above it — deployment containers, agent harnesses, inference integrations, customization pipelines — are open. Vercel appears on the launch platform list alongside AWS, Google Cloud, Azure, Cloudflare, CoreWeave, Together AI, Fireworks, OpenRouter, Baseten, Modal, NVIDIA NIMs, Ollama, and LM Studio.

The company is also open-sourcing Mirror, a CLI coding-agent harness it built for internal use. Mirror pairs a Python product layer with a Rust agent core, supports resumable conversations and parallel agents in separate worktrees, and exposes inspection and tool-call approval. Beam has also been tested with Pi, OpenCode, and Hermes.

On the training side, Reflection reports roughly 1.3 billion sandboxes across Beam's reinforcement-learning runs, with the fleet peaking at 883,006 vCPUs across 23 clusters on GCP and AWS. Fine-tuning and RL workflows are supported through partnerships with NVIDIA (NeMo RL, AutoModel, Megatron), Hugging Face TRL, Anyscale SkyRL, RadixArk Miles, and Axolotl. Self-distillation recipes — generate traces on your tasks, train a LoRA adapter on verified traces, repeat — are promised in the coming days.

Should you swap your coding agent?

Apache 2.0 means commercial use, modification, and redistribution without license fees. Self-hosting the full model needs serious GPU capacity even with 23B active parameters, so most developers will start with a hosted endpoint. A fair test: run reflection/beam-free against a real repo task and compare it with your current default before committing anything. If the free window survives, it's the cheapest way to evaluate a 501B-class open model right now.

Beam is one of the most significant open-weight coding releases to date, and Reflection clearly intends it as the base layer of an open agent stack rather than a one-off drop. What remains unproven is independent benchmark performance — the official model card numbers are vendor-reported, not third-party verified. The next milestone to watch: those self-distillation recipes and how quickly the free window converts to paid pricing.

For building agents on this kind of model, see our guide to AI agent tools.