Reflection AI Unveils Beam Open Weight Model with 501B Sparse MoE Architecture

Sparse Mixture of Experts Architecture

Reflection AI announced its first open-weight model named Beam in early October 2026. Founded by former Google DeepMind researchers and backed by Nvidia, the startup engineered Beam specifically for software engineering, rigorous reasoning, and autonomous agent workloads.

The model utilizes a sparse Mixture of Experts (MoE) configuration with the following technical specifications:

• Total parameters: 501 billion parameters across the entire network.

• Active parameters: 23 billion parameters activated per processed token.

• Context window: Supports up to 1,000,000 tokens in a single session.

• Inference efficiency: Consumes three to four times less inference compute than comparable proprietary models on advanced reasoning benchmarks.

By limiting active parameter activation per forward pass, Beam provides high generation quality while reducing memory and hardware overhead for enterprise deployments.

Pretraining Pipeline and Compute Infrastructure

Training the model required a massive allocation of accelerator capacity:

• Pretraining volume: The base model ingested 23.8 trillion tokens spanning natural language, technical documentation, and source code.

• Reinforcement learning phase: Over 100 million reinforcement learning rollouts were executed to train autonomous task validation and code syntax checks.

• Hardware cluster: Training completed across a four-week run powered by 10,500 Nvidia GB300 GPUs.

Internal evaluations place Beam competitively against top reasoning architectures such as GLM-5.2 on programming challenges and multi-step tool use benchmarks.

Licensing Framework and Weight Availability

Reflection AI is adopting an open distribution model to address the scarcity of high-tier open-weight architectures outside closed API endpoints:

• Apache 2.0 license: Model weights, a detailed technical report, and system evaluation cards will be released publicly under an Apache 2.0 license later in October 2026.

• Pre-release testing: The model is currently undergoing red teaming and private waitlist evaluations with enterprise partners.

• Self-hosted autonomy: The open license permits organizations to host, fine-tune, and inspect model layers on private infrastructure without external network dependencies.

This release provides developers and institutions with a fully auditable frontier model suitable for secure, air-gapped agent operations.