Beam Operational

Beam is Reflection AI's first open-source large language model. It uses a 501B-parameter sparse MoE architecture, activates only 23B parameters per token, supports a 1-million-token context window, and was pretrained on 23.8T tokens. Beam's training process represents an industrial-scale showcase of computing power: pretraining was completed in under four weeks on 6,144 GB300 GPUs, followed by four weeks of large-scale reinforcement learning on 10,500 GB300 GPUs with more than 100 million rollouts.

Software development: Reaches a level comparable to GLM-5.2 on software engineering tasks and serves as a primary model for enterprise coding workloads.
Agent execution: Supports tool calling, terminal operations, and web search. It can complete multi-step tasks autonomously and has even learned to call other large models and OCR APIs.
Long-context processing: Supports a 1-million-token context window, enabling it to locate and connect key information across extremely large codebases and documents.
Adjustable reasoning effort: Users can switch flexibly between short responses and extended reasoning through the reasoning effort parameter, matching different tasks and compute budgets.

Follow WeChat and reply “开源” to join the AI open-source project community.
Sparse MoE architecture: The model has 501B total parameters, but activates only 23B at a time. Fine-grained expert routing, together with auxiliary-loss-free load balancing, achieves nearly uniform expert utilization. The busiest expert reaches only 1.04× the average load, providing the hardware-level foundation for Beam's efficient inference.
Two-stage large-scale training: Pretraining consumed 23.8T tokens in under four weeks on 6,144 GB300 GPUs. With a topology-aware scheduler and silent data-corruption detection, it achieved 92.3% goodput. This was followed by four weeks of reinforcement learning on 10.5K GB300 GPUs, producing more than 100 million rollouts while maintaining an average of 110,000 concurrent rollouts and 170,000 concurrent sandboxes.
Fully asynchronous RL system: Asynchronous policy gradients enable rollout and training to run in parallel. Each token is tagged with the model version that generated it to address policy staleness. Numerical stability is maintained even when samples lag one day behind the current policy (107 versions). New weights are distributed hierarchically to inference clusters over RoCE in a median of just 12 seconds.
Length-controlled reasoning learning: A length penalty is added during RL so the model first learns to solve problems with fewer tokens, then naturally expands its reasoning length as agentic capabilities improve. This creates an optimal tradeoff between capability and token consumption that users can adjust as needed.

The weights have not yet been released. A preview version is currently available only to selected users. Visit https://platform.reflection.ai/ to apply for early access.

Extremely high inference efficiency: Achieves reasoning performance at the same level as GLM-5.2 with 3–4 times less inference compute. It saves more compute than models such as Qwen 3.8-Max with 2T+ parameters, delivering outstanding intelligence density.
Minimal active parameters: Only 23B of the 501B total parameters are activated per token, and the sparse MoE design substantially reduces deployment and inference costs.
Exceptional RL scale: Four continuous weeks on 10.5K GB300 GPUs and more than 100 million rollouts make this one of the largest RL training runs to date among open laboratories.
Strong training stability: Pretraining required only 9 semi-automatic rollbacks and achieved 92.3% goodput, with no instability in the training curves, demonstrating a high level of engineering maturity.
Controllable reasoning length: Users can freely adjust reasoning depth through the reasoning effort parameter, balancing performance and cost as needed.

Official website: https://reflection.ai/blog/introducing-beam

Private enterprise coding assistants: Enterprises can fully self-host low-cost AI coding assistants while keeping data within their internal networks.
Codebase-scale refactoring and auditing: With its 1-million-token context window, Beam can load an entire medium-sized codebase at once, understand dependencies across files, and perform large-scale refactoring or security audits.
Automated agent workflows: After reinforcement learning with more than 100 million rollouts, Beam has developed tool-calling, terminal-operation, and web-search capabilities. It can drive long-chain, multi-step tasks such as data analysis and operations inspections.
Sovereign AI deployment for government and enterprises: Developed by a U.S.-based team as an open-weights model, Beam can be audited and deployed in isolated environments, meeting the compliant procurement needs of government, energy, and defense organizations seeking models outside the Chinese supply chain.