Reflection AI introduced Beam on October 5, 2026 as its first open-weight frontier model, targeting coding, reasoning, and long-running agentic workloads.
What changed
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters. Reflection says it was pretrained on 23.8 trillion tokens and strengthened through large-scale reinforcement learning. The company says its training used more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks.
Why the architecture matters
A sparse model activates only part of its network for each token, which can reduce inference computation compared with a dense model of similar total size. For developers, the important metric is therefore compute and cost per completed task rather than parameter count alone.
Built for coding and agents
Reflection is positioning Beam for software engineering and agentic workflows. The company reports a one-million-token context window and says the model is intended to compete with leading open models on coding and reasoning tasks.
Open-weight strategy
Reflection plans to release model weights and full technical details during October 2026. Open weights can give enterprises and public-sector teams more control over deployment, data, customization, and infrastructure choices.
Important caveat
Performance and efficiency figures announced at launch are company-reported and should be independently tested when the weights and technical report are available.
Practical takeaway: Beam is significant because it combines frontier-scale total parameters with sparse activation and an open-weight deployment model. Developers should evaluate cost per completed coding or agent task rather than relying only on headline benchmark scores.
Source: Reflection AI, Introducing Beam, October 5, 2026.