Beam 1.0 501B A23B
Built for coding, multi-step tool use, and long-running agents that need the capacity of a large open model at the per-token compute cost of a much smaller one.
Explore additional models
Frequently
asked questions
Beam 1.0 501B A23B is Reflection AI’s first open-weight model, a mixture-of-experts style model with 501 billion total parameters but only 23 billion active at any one time. It is text-only, supports a 1M-token context, and reasons before every answer with adjustable effort. Reflection AI trained it for coding, tool use, and long-running agent work, and released it under the Apache 2.0 license in October 2026. It is available for high-performance inference via Self-Serve Deployments in Crusoe Intelligence Foundry.
The “A23B” stands for about 23 billion active parameters. Beam holds 501B parameters across its experts but routes each token through only about 23B of them, so each token costs compute closer to that of a 23B model while the model keeps the knowledge of a much larger one. All 501B parameters still load into GPU memory, so the savings show up in latency and throughput, not memory footprint.
Reflection AI built Beam 1.0 501B A23B for coding work, including agentic coding in real codebases and multi-step tasks in the terminal, as well as scientific reasoning. It’s a mixture-of-experts model with 501 billion total parameters and about 23B active per token, and Reflection AI publishes it with BF16 weights.
Yes. Beam 1.0 501B A23B is available for high-performance inference via Self-Serve Deployments, with profiles tuned for throughput or responsiveness.


%204.png)
