TL;DR: Reflection AI launched Beam 1.0, a 501B parameter Mixture-of-Experts (MoE) model optimized for NVIDIA Blackwell GPUs, achieving over 70,000 tokens/second throughput.
Summary: Reflection AI has released Beam 1.0, a 501B parameter Mixture-of-Experts (MoE) model. This model is specifically optimized for NVIDIA Blackwell GPUs, achieving a throughput exceeding 70,000 tokens/second on the NVIDIA GB300 NVL72 platform. It offers multiple weight formats (BF16, FP8, NVFP4) and includes an open-source framework for deployment and evaluation.
Why it matters: This release demonstrates significant advancements in model efficiency and scalability for large language models. AI builders should explore Beam 1.0 for high-throughput inference on NVIDIA Blackwell hardware, especially for applications requiring rapid token generation.
Source: rss