Test-time Compute is the practice of spending more compute at inference to improve accuracy. This guide covers OpenAI o1/o3, DeepSeek-R1, the Chain-of-Thought connection, and how production systems balance latency, cost, and quality.

Mamba is a Selective State Space Model architecture introduced by Gu and Dao in 2023 that scales linearly with sequence length, addressing the Transformer's quadratic-attention bottleneck. Hybrid Mamba-Attention models like Qwen3.6 and Jamba ship in production today.

Speculative Decoding accelerates LLM inference 2-3x without changing the output distribution. A small draft model proposes tokens, the target model verifies them in a single forward pass, and only matching tokens are accepted. Standard in vLLM, TGI, and TensorRT-LLM.

FlashAttention is Tri Dao's GPU-memory-aware attention algorithm delivering 2-4x speedup and long-context training. FA-3 reaches 75% of H100 FLOPS. Architecture, patterns, comparison.

vLLM is an open-source LLM serving engine using PagedAttention for near-100% GPU memory utilization and up to 24x throughput. Architecture, patterns, and TGI/TensorRT-LLM comparison.

The KV Cache is the optimization that turns Transformer LLM inference from quadratic to linear by reusing computed Key and Value tensors. Learn how it works, its memory cost, and how it relates to Prompt Caching, PagedAttention, and vLLM.

Quantization compresses LLM weights into lower-precision formats to save memory and speed up inference. This guide covers FP16/INT8/INT4, GPTQ, AWQ, GGUF, hallucination trade-offs, and production patterns.

Diffusion Model is a class of generative AI that produces images, audio, and video by gradually denoising random noise. This guide covers the math intuitively, how Stable Diffusion and Sora 2 use it, and how it compares to GANs and autoregressive models.

A complete guide to Mixture of Experts (MoE), the sparse Transformer architecture that powers Mixtral, DeepSeek V3, and Llama 4. Learn how routing works, total vs active parameters, training challenges, and dense-vs-sparse trade-offs.

A reasoning model is a large language model that internally produces an explicit chain of thought before answering. Learn how OpenAI o3, Claude Extended Thinking, Gemini 2.5 Thinking, and DeepSeek R1 work, their differences from standard LLMs, and when to use them.