What is NVIDIA-Backed Reflection AI & its Open-Weight Model?
We unpack NVIDIA-backed Reflection AI, its US open-weight strategy, the NVIDIA Nemotron Coalition and how its Beam model maximises efficiency
TL;DR
- Reflection AI is a US startup focused on building open-weight frontier AI models.
- The company was founded by former Google DeepMind Staff Research Scientists Misha Laskin and Ioannis Antonoglou.
- Reflection AI aims to be a counterweight to Chinese AI labs, which often have open-weight or open-source models.
- Their latest model, Beam, has 501 billion parameters and is designed for inference efficiency, activating only 23 billion parameters per token.
- Beam is trained on 23.8 trillion tokens and will be open-sourced under an Apache 2.0 license.
- Reflection AI is part of the NVIDIA Nemotron Coalition, a collaboration focused on advancing frontier open models.
- The startup is reportedly in talks to raise $2.5 billion at a $25 billion valuation.
- Partnerships include DELL, the US Department of Energy, and SHINSEGAE.
- Misha Laskin emphasizes the importance of US alternatives to open-weight models for infrastructure and preventing dependence on foreign technology.