Reflection AI announced Beam, a mixture of experts with 501 billion parameters and 23 billion active per token. It was pretrained on 23.8 trillion tokens in under four weeks, on 6,144 Nvidia GB300 GPUs, with a 1M token context. Reflection reports 80.9 on SWE-bench Verified, above Inkling at 77.6. On Humanity's Last Exam it trails GLM 5.3, Qwen 3.8-Max and Kimi K3. The weights are not out. Early access is a waitlist, with Apache 2.0 reportedly planned this month. Every figure is the company's own.