Understand the numbers.

Short guides to the quantities that decide whether a model fits, how fast it runs, and which runtime makes sense.

Bits per weightWhat one parameter costs, and why the number is never round.Why bandwidth decides speedThe memory bus finishes last, so it sets the pace.KV cache and contextWhy a long conversation is slower than a short one on the same machine.MoE vs dense modelsTotal parameters decide the fit, active parameters decide the speed.OffloadingWhat happens when the weights do not all fit, and the one exception worth taking.Prefill vs decodeWhy one machine reports two very different token rates.MTP and speculative decodingDraft tokens buy real speed, and the multipliers do not stack.Choosing an inference stackThe runtime can move the answer more than any tuning knob.Quantization formatsChoosing among Q8, Q4, activation-aware, and native low-bit.Scale and zero pointThe arithmetic under every quant format: a float becomes an integer, and back.GGUF quant familiesWhat _0, _K, IQ, and the S/M/L suffixes mean in a filename.Low-bit trainingWhy 1.58-bit models exist, and why you cannot get there after the fact.Multi-GPU scalingMemory adds up cleanly. Bandwidth depends on the interconnect.Uncertainty and calibrationWhy the answer is a band, and when to stop trusting it.The craft of llm.fitHow the interface keeps estimates specific, inspectable, and honest.