Reflection released Beam’s model weights on October 16, making the 501-billion-parameter model available to download and customize under the Apache 2.0 license. The release follows the company’s October 5 announcement. But Reflection’s published scores still put leading Chinese models ahead on Terminal-Bench 2.1.

Developers can now test and adapt Beam themselves, but its benchmark results have not established a lead over top Chinese models.

What developers can access

Beam is a mixture-of-experts model. Its model card lists about 501 billion parameters in total, with 23 billion active for each token.

  • Available weight formats: BF16, FP8 and NVFP4
  • Included materials: a model card and safety report
  • License: Apache 2.0

Developers can run Beam on their own infrastructure, adapt it for specific tasks and connect it to agent systems. But the license does not remove the hardware demands: self-hosting a model of this size still requires substantial computing resources.

Efficiency is not the same as lower cost

Reflection says Beam can reach results comparable to leading open models while using fewer reasoning tokens. That could reduce the tokens needed for a task, but it does not prove that a deployment will cost less. Hardware, software and provider pricing also affect the bill.

VentureBeat’s analysis noted that Reflection’s published Terminal-Bench 2.1 scores put leading Chinese models ahead of Beam. The benchmark measures agent performance on command-line tasks. It offers one measure of capability, not a verdict on every coding or reasoning use.

For developers, the immediate change is access: Beam can now be tested and adapted directly, rather than judged only through company claims or hosted access. Whether its token efficiency offsets the cost of running a model this large will depend on the workload and deployment.