Reflection opens Beam weights: 501B-parameter sparse MoE, 23B active, Apache 2.0 self-hosting
Synopsis
Reflection released the weights of Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license; Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning, and long-horizon agentic workloads. Previously the model was reachable only through Reflection and its partners, and developers can now download and self-host it, evaluate it, and customize it for their own domains. The release ships with final benchmarks in the model card and a safety technical report, plus an ecosystem of over 40 partners and integrations spanning inference, quantization, fine-tuning, reinforcement learning, agent harnesses, evaluation, and observability, which Reflection calls an open operating system for agents.
Interpretation
Beam's weights can now be downloaded and self-hosted, so developers can evaluate the model, customize it for their own domains, and connect it to agentic systems. Previously Beam was reachable only through Reflection and its partners, with no public weights. The weights are released on Hugging Face under an Apache 2.0 license, with final benchmarks in the model card and a safety technical report covering the complete safety methodology and evaluations.
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning, and long-horizon agentic workloads. Sparse Mixture-of-Experts means each forward pass calls only part of the expert sub-networks, so active parameters are far fewer than total parameters. Reflection describes it as built for frontier reasoning efficiency and reports seeing use cases from early developers.
The weights ship in BF16, FP8, and NVFP4 precision formats along with a speculative decoding drafter model, and partners optimized serving configurations and developed fused kernels for lower latency and higher throughput. Together these reduce memory requirements and accelerate token generation out of the box. Beam has native day-zero support across Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and Dynamo, so developers can go from weights to a production-scale endpoint without writing custom plumbing.
Perspective
This report is for researchers and engineers following open-weight model releases, and it covers the release of Beam's weights, documentation, and companion ecosystem itself. The material is a summary and does not give specific benchmark numbers, safety evaluation results, or details of the early developer use cases. Reflection states its vision is to build an open intelligence stack that gives developers, enterprises, and governments control over how AI is deployed in their environments.
The material is a summary and does not list the final benchmark numbers from the model card or the specific evaluation results of the safety technical report. Reflection says it has seen use cases from early developers but does not describe them in this material. Beam was trained alongside Mirror CLI for best performance while also integrating with other open source harnesses.