The field of large language model (LLM) inference is moving towards ultra-low-bit quantization, which enables significant reductions in computational costs and memory requirements. Recent developments have introduced innovative quantization methods, such as 2-bit quantization, that provide improved information density and deterministic gradient information. Additionally, there is a growing focus on optimizing LLM inference for resource-constrained environments, such as edge devices and AI PCs, with advances in microkernel design and runtime optimization. Noteworthy papers in this area include: The Fourth State: Signed-Zero Ternary for Stable LLM Quantization, which introduces a 2-bit quantization method that improves information density. Pushing the Envelope of LLM Inference on AI-PC, which presents a state-of-the-art LLM inference framework that achieves up to 2.2x better performance than existing runtimes. Rethinking 1-bit Optimization Leveraging Pre-trained Large Language Models, which proposes a consistent progressive training method for 1-bit LLM quantization that outperforms existing approaches. Profiling Large Language Model Inference on Apple Silicon, which investigates the efficiency of Apple Silicon for on-device LLM inference and debunks existing myths about its performance. XQuant: Breaking the Memory Wall for LLM Inference with KV Cache Rematerialization, which presents a novel quantization method that reduces memory consumption by an order of magnitude with substantial accuracy benefits.