Reflection Releases Beam: 501B-Parameter Sparse MoE Model Under Apache 2.0 License
Reflection releases Beam, an open-weight 501B sparse MoE model with 23B active parameters, 1M context window, and Apache 2.0 licensing for agentic and coding workloads.
Reflection has officially released the weights for Beam, a 501-billion-parameter sparse Mixture-of-Experts (MoE) foundation model under the permissive Apache 2.0 license. Featuring 23 billion active parameters per token, a 1-million-token context window, and pretraining across 23.8 trillion tokens, Beam is engineered specifically for reasoning, complex coding, and autonomous agent workflows at 3-4x lower inference compute.
Sparse Mixture-of-Experts Architecture and Extreme Inference Efficiency
Announced on October 18, 2026, Beam utilizes a sparse MoE architecture comprising 52 layers with 256 experts per layer, dynamically routing tokens through only 8 experts during inference. This design yields 23 billion active parameters per token from its 501-billion-parameter total pool, providing frontier-tier reasoning while dramatically slashing compute overhead. Reflection reports that Beam achieves 3-4x lower inference compute than leading dense open models, supported by high-compute reinforcement learning trained on 10,500 NVIDIA GB300 GPUs across more than 100 million rollouts.
Frontier Benchmark Performance and 1M-Token Context Window
Trained on 23.8 trillion tokens of curated pretraining data, Beam features a massive 1-million-token context window capable of ingesting entire multi-repository codebases in a single prompt. Internal benchmarks place BF16 Beam at 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1, 90.6% on GPQA Diamond, and 97.7% on AIME 2026, positioning the text-focused model as a premier engine for long-horizon software engineering and technical problem solving.
Permissive Apache 2.0 Licensing and Broad Ecosystem Deployment
Published under the unrestrictive Apache 2.0 license on Hugging Face, Beam allows unrestricted commercial adoption, distillation, and fine-tuning. The model launches with weights in BF16, FP8, and an NVFP4 quantized variant that compresses 1TB of parameters down to 280GB to fit four B200 accelerators. With day-zero support across vLLM, SGLang, Transformers, and hosted inference endpoints on AWS, Azure, Google Cloud, and CoreWeave, Reflection establishes Beam as a cornerstone open-source operating system for autonomous agents.