On October 16, 2026, Reflection published the model weights of Beam on Hugging Face under an Apache 2.0 license. Reflection calls Beam its first open-weight model, “built for advanced coding, reasoning, and long-horizon agentic workloads.” It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active. Reflection released a model card with final benchmarks and a Beam Safety Technical Report alongside the weights.
The weights ship in BF16, FP8 and NVFP4 precision, along with a speculative-decoding drafter model. Reflection lists day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, and says more than 40 partners and integrations launched with it, which it calls “an open operating system for agents.”
Reflection also presented Mirror CLI, a terminal coding and agent harness it says Beam was trained alongside and which it has evolved into a tool it could open source. Mirror has a Python front end and a Rust agent core, supports resumable conversations and parallel agents in separate worktrees, and accepts skills, MCP connections and Python extensions. Beam has also been tested with other open harnesses including OpenCode and Hermes.
Why it matters: Beam is a frontier-scale model from a US lab released under a fully permissive Apache 2.0 license, and it arrives with its own agent harness, multiple precision formats and day-zero serving support rather than as bare weights, which is what makes an open model practical for enterprises and governments to run themselves.
What it does not show: the benchmark results are Reflection’s own, published in its model card, and this page does not state them. Independent evaluations had not been published at release, and the earlier API preview announcement is not covered here.