Reflection opens 501B-parameter Beam weights under Apache 2.0, with 40-plus partners turning one model into a self-hostable stack

An open-weight release now arrives as more than weight files: inference, quantization, fine-tuning, reinforcement learning, agent runtimes and deployment paths come with it, a combination Reflection calls an open operating system for agents in which the model is the kernel and everything above the infrastructure layer is open. Previously a single open-weight release covered the model layer alone, and making it useful still required reliable inference, evaluation, safety tooling, customization workflows and deployment channels. Reflection says it worked with open-ecosystem partners on inference, quantization, fine-tuning, reinforcement learning, agent harnesses, evaluation and observability, with partners offering Beam through managed endpoints, cloud platforms, on-premises systems and sovereign infrastructure.

Beam uses a sparse Mixture-of-Experts design with 501 billion total parameters and 23 billion active, aimed at coding, reasoning and long-horizon agentic workloads with frontier reasoning efficiency. Open-weight models have generally arrived as dense designs or at smaller scale, so separating active from total parameters is uncommon on the open-weight side. Reflection states this parameter configuration in the release and says the model card carries final benchmarks while the Beam Safety Technical Report shares the complete safety methodology and evaluations.

Weights ship in BF16, FP8 and NVFP4 precision formats with a speculative decoding drafter model, and partner-optimized serving configurations and fused kernels cut memory requirements and accelerate token generation out of the box. Previously developers who obtained weights often had to handle precision conversion and serving tuning themselves before reaching a production-scale endpoint. Reflection says Beam has native day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and Dynamo, so users can go from weights to a production-scale endpoint without writing custom plumbing.

Reflection also open-sourced Mirror CLI, the harness it built for internal use, with a Python frontend for prompts, skills and developer tooling talking over a local connection to a Rust agent core that drives the model, kept minimal so it doubles as a clean evaluation surface. Coding agent tools have often been tied to a specific model or closed service, leaving developers little room to point them at different checkpoints or endpoints. Reflection says Mirror supports resumable conversations and parallel agents in separate worktrees, and that Beam was tested with Pi, OpenCode and Hermes across coding and knowledge work including reading repositories, editing files and running commands.

This report covers Reflection's release of Beam open weights, documentation and partner ecosystem, along with the parameter configuration, precision formats, supported frameworks and deployment channels stated in the release. It is intended for researchers and engineers following open-weight models, agent tooling and self-hosted deployment. The release says self-distillation recipes will follow in the coming days and that lessons from training Beam are being upstreamed into open standards such as OpenEnv for future runs.

The release says the model card carries final benchmarks, but no concrete numbers appear in the text, so capability comparisons cannot be made from this material alone. Reflection says self-distillation recipes will be released in the coming days, meaning that customization path is not yet available with the weights. The release also says harness support will keep expanding, so the current framework coverage is expected to change.