US startup Reflection announced Beam on October 5. The 501-billion-parameter model is designed for coding, reasoning and AI-agent tasks, but only select preview users can access it. Developers cannot yet download its weights.

That makes Beam’s performance claims difficult for outside developers to check. Reflection says it plans to release the weights and documentation later in October, after final evaluations.

What 501 billion means

Beam uses a sparse mixture-of-experts design. Reflection says the model has 501 billion parameters in total, but uses 23 billion for each task.

  • 501 billion total parameters
  • 23 billion active for each task, according to Reflection

The difference describes how the model uses its parameters. It does not, by itself, show how quickly or cheaply Beam will run in practice.

Reflection says Beam is competitive with Z.ai’s GLM-5.2 and approaches Alibaba’s Qwen 3.8-Max on some coding and agentic tasks. The company cites selected benchmark results. Those tests cover specific tasks and conditions, so they cannot establish how the model will perform across everyday development work or against every competing system.

Public access is the next test

Beam is currently in a limited early-access preview. Its promised public release of weights, technical documentation and related materials has not happened yet.

Reflection says it is conducting final evaluations and expects to publish those materials later in October. Semafor reported that the startup is pitching Beam to businesses and governments seeking more control over how they use AI.

For developers, the practical question is whether the promised release arrives and gives them a chance to run Beam themselves. Until then, its performance comparisons remain claims from Reflection, not results confirmed through public testing.