Reflection AI’s Beam pairs open weights with a million-token context

Reflection AI has released Beam, an open-weight model aimed at coding and agentic tasks, with a one-million-token context window. Its Apache 2.0 licence and availability across several inference platforms make this a substantial release for developers deciding what to run and where.

Reflection AI Watch analysis

What happened

Reflection AI’s model card describes Beam as a 501-billion-parameter mixture-of-experts model, with 23 billion parameters active at a time. It says the model is available through Hugging Face, OpenRouter, Baseten, Vercel AI Gateway and CoreWeave Serverless Inference.

The Beam model card also outlines its training data and processing, including deduplication and filtering intended to reduce personal information and harmful content.

Our top picks

• Open weights under Apache 2.0
  Developers can inspect and use the released weights under a permissive licence.
• A one-million-token context window
  Beam is designed to handle unusually large amounts of context in one go.
• 23 billion active parameters
  Its mixture-of-experts design activates a fraction of its 501 billion total parameters for a given task.
• Reasoning-effort controls
  Users can adjust the reasoning budget rather than treating every prompt as an all-out effort.
• Coding and agentic focus
  Reflection positions Beam for coding work and tasks involving tool use.
• Multiple ways to run it
  The model card lists availability through five platforms, including Hugging Face and CoreWeave.

Why it matters

Beam combines open weights with a very large context window and a focus on coding and tool use. That gives developers several practical choices: run the model themselves, adjust its reasoning effort, or use an inference provider. The one-million-token figure is an advertised context length, not by itself evidence of how well the model uses every token.

The card also cites provider-specific performance claims: Google Cloud says its setup can deliver up to five times higher performance, while Applied Compute says its reinforcement-learning post-training tripled performance on long-horizon coding tasks. Those are claims about particular systems and evaluations, not a universal speed or quality guarantee.

Our read

Beam has the ingredients of a meaningful open-model release, especially for teams that want a long context and control over where inference happens. The useful next step is to compare it on your own workloads, rather than letting a striking parameter count do the benchmarking for you.

What to watch

• Independent evaluations of coding, tool use and long-context performance.
• The hardware and serving costs of running a 501-billion-parameter model.
• Whether the provider-specific performance gains hold across workloads.

Discussion spark: For an open model aimed at coding agents, which matters more to you: a million-token context window, controllable reasoning effort, or performance on independent tests?

not affiliated with or endorsed by Reflection AI