Model release · Reflection AI
Reflection AI releases open-weight Beam, a 501B MoE, with NVIDIA NVFP4 support
- Released
- Review updated
Reflection AI released Beam, its first open-weight model, under the Apache 2.0 license. The 501B-parameter MoE was trained on NVIDIA GB300 NVL72. The release includes weights, a tokenizer and a reference implementation, with BF16, FP8 and NVFP4 weights plus a DFlash2 draft model. NVIDIA published a Beam 1.0 NVFP4 checkpoint on Hugging Face, reporting throughput above 70,000 tokens per second per GPU on Blackwell Ultra; a later update raised that figure to about 90,000. These are vendor-reported throughput figures for the stated hardware, not single-request speeds or guarantees for other systems.
Details
Reflection AI released Beam, its first open-weight model, under the Apache 2.0 license. The 501B-parameter MoE was trained on NVIDIA GB300 NVL72. The release includes weights, a tokenizer and a reference implementation, with BF16, FP8 and NVFP4 weights plus a DFlash2 draft model. NVIDIA announced its Beam 1.0 NVFP4 checkpoint at 21:39 Beijing time on October 16, 2026, reporting throughput above 70,000 tokens per second per GPU on Blackwell Ultra; a later update raised that figure to about 90,000. These are vendor-reported throughput figures for the stated hardware, not single-request speeds or guarantees for other systems.