Reflection AI Releases Beam Model Weights on Hugging Face

Reflection AI announced that its Beam large‑language model, a 501‑billion‑parameter mixture‑of‑experts network, is now available on Hugging Face under an Apache 2.0 license. The release includes the full weights in BF16, FP8 and NVFP4 precision formats, a lightweight drafter model that supports speculative decoding, and a suite of post‑training recipes and code templates that let developers fine‑tune Beam for specific tasks. Beam was pretrained on 23.8 trillion tokens and then refined through more than 100 million reinforcement‑learning rollouts on 10.5K NVIDIA GB300 GPUs over four weeks, and the company supplied a model card and a comprehensive safety technical report.

Beam’s launch is backed by an ecosystem of more than 40 partners that provide day‑zero support and deployment containers. Baseten, Amazon Web Services, Microsoft Azure, Google Cloud and others now host the model, with Baseten offering a free‑to‑start API and per‑token pricing. Developers can self‑host Beam, evaluate its performance, customize it for domain‑specific workloads, and connect it to agentic systems through Reflection’s Mirror CLI, a terminal agent harness that supports resumable conversations and parallel agents. The company also plans to release self‑distillation recipes in the coming days and to adopt the OpenEnv standard for future training runs.