Efficient inference is becoming an architectural frontier

TokenRouter reports 2.01–64.15x decoding-throughput gains for token-level LLM routing; SparseDecoding adds up to 1.48x end-to-end speedup through decoding-aware pruning and N:M kernels, while V-CoLA reports 1.86–6.15x multimodal prefill gains with 99.5% performance retention at half the vision tokens. Quantization evidence favors 4-bit inference, but realistic hardware, workload, and quality validation remain the key test.