AI developer Reflection AI has officially announced Beam, marking the team's first open-weight model release. The launch represents a strategic move to establish competitive frontier-grade open models within the United States. Beam is built on a specialized sparse mixture-of-experts architecture with a total parameter count of 501 billion. The model weights are scheduled for public release under the permissive Apache 2.0 license during October 2026.
The architectural centerpiece of Beam is its aggressive sparsity during inference. While the complete parameter count reaches 501 billion, only 23 billion parameters are actively routed per processed token. This represents an activation rate of approximately 4.6 percent. By keeping active weights low, Beam delivers fast execution and reduces operational computational demands significantly compared to dense models.
Training for the system encompassed an extensive corpus of 23.8 trillion tokens. Reflection AI specifically tailored the training distribution to excel at software engineering workflows and persistent agent tasks. The system was designed from the ground up to operate reliably across complex terminal sessions and development environments. This specialized focus is reflected directly in the initial evaluation metrics published by the team.
On standard industry benchmarks, Beam demonstrated notable strength across rigorous programming evaluations. The model scored 80.9 points on the SWE-bench Verified benchmark according to Reflection AI. Furthermore, it achieved 80.1 points on Terminal Bench, highlighting its capacity to handle multi-step system tasks and shell operations. These numbers position Beam in direct competition with prominent frontier coding systems.
Strategically, Reflection AI positions Beam as an American open-weights alternative to prominent Chinese models such as GLM 5.2 and the Qwen series. Over recent release cycles, open-weights momentum has heavily centered on international labs. Beam intends to match high reasoning performance while cutting inference compute requirements by a factor of three to four. This balance makes scalable agent deployment far more viable for enterprise developers.
The forthcoming Apache 2.0 release grants teams complete freedom to deploy, fine-tune and inspect the model in self-hosted environments. Enterprises handling proprietary codebases can leverage advanced coding automation without transmitting intellectual property to closed external platforms. With only 23 billion active parameters needing compute during generation, the operational hardware barrier drops substantially. The release marks a meaningful step forward for open-weights agent infrastructure.