
Reflection Beam Throughput and VRAM: What 501B MoE Inference Costs Per Token
pr0h0•
A practical look at what Reflection AI's 501B-parameter Beam MoE, with only 23B active parameters per token, actually costs in VRAM, throughput, and dollars per token when self-hosted for agentic coding workloads.
mixture-of-expertsllm-inferencevram