Reflection AI open-sources Beam’s weights, post-training recipes

Reflection AI Inc. has made the weights of its first large language model, Beam, available for open-source use on Hugging Face. The company also released technical resources to help developers customize Beam. Beam, introduced last week, is one of the most advanced open-source LLMs, boasting 500 billion parameters but matching the performance of larger models with less computing power and memory.

Training Beam, which took two months using several thousand high-end graphics cards, involved significant error mitigation efforts. Reflection AI developed custom reliability optimization algorithms to address potential errors during training-inference mismatches. The company also created a system to automatically fix data quality errors, allowing for training with minimal cluster downtime.

Beam's weights come in standard and scaled-down versions using quantization, balancing quality and hardware efficiency. Alongside these weights, developers can access training recipes, fine-tuning code templates, and measurement tools to assess customization impacts and safety. Reflection AI also unveiled a draft model for Beam, facilitating speculative decoding to reduce hardware usage.

This approach involves a lightweight LLM generating an initial response prompt, which Beam then checks for errors before user display. The release includes pre-configured Beam settings for popular serving engines and fused kernels to expedite processing. Reflection AI plans to follow up with additional training recipes enabling self-distillation, enhancing output quality. The company intends to contribute insights from Beam development to open-source AI tools.