Reflection releases Beam 1.0, a 501B open MoE

Reflection has published Beam 1.0, a 501-billion-parameter mixture-of-experts model released under the permissive Apache 2.0 license. The weights are available on Hugging Face, and the company lays out its reasoning behind the launch in an announcement post.

The headline design choice is sparsity. While the full model carries 501B parameters, only about 23B are active for any given token. That architecture lets Beam reach the capacity of a very large model while keeping per-token compute closer to a mid-sized dense system — a tradeoff that has become the default for ambitious open releases.

What it's built for

Reflection positions Beam squarely as a reasoning model, with additional emphasis on code generation and tool use. Those are the workloads where extended, multi-step inference tends to pay off, and where sparse MoE designs have been competitive with far larger dense alternatives.

• 501B total parameters, roughly 23B active per token
• Apache 2.0 license, permitting commercial use
• Focus on reasoning, code, and tool-calling

Why it matters: Reflection is best known as a well-funded startup pursuing frontier capabilities, and an Apache-licensed MoE at this scale signals a serious commitment to the open-weights ecosystem rather than a token gesture. For teams building agentic or code-heavy systems, a permissively licensed model with large capacity and modest active compute is a meaningful option — though real-world evaluation against established open MoEs will determine where Beam lands.