Matryoshka LM Suites and Training Efficiency Innovations

New paper introduces Matryoshka LM Suites, nesting multiple models in one architecture to train a suite in a single run, reducing compute while maintaining performance, with free distillation and better speculative decoding. Complements other efficiency advances like FreeToken (753B MoE on single workstation), Pathway's cost-efficient reasoning, and Liquid AI's DSpark draft models (up to 3.18x faster decoding, 57% latency reduction for agentic workflows). Highlights a trend toward compute-efficient training and inference.