Reflection AI has made Beam's weights available for download, giving developers a way to customize the model and evaluate resulting changes to its capabilities and safety.
Reflection's model card puts Beam's context limit at 1 million tokens. It lists 501.3 billion parameters overall, including 23.4 billion active.
Beam's text weights come in BF16, FP8 and NVFP4 formats. Reflection lists Apache 2.0 as their license.
Partners have supplied fine-tuning and reinforcement-learning workflows, alongside tools for assessing changes to Beam's capabilities and safety. Reflection also released a drafter model for speculative decoding. The company plans to publish self-distillation recipes in the coming days.
Reflection's Oct. 16 announcement dates Beam's unveiling to Oct. 5 and the weight release to Oct. 16. The company said the launch ecosystem included more than 40 partners and integrations.
