Concepts

Task FitIs this model right for the job?Model CompareTwo to four models side by side Model AnatomyWhere its parameters and memory go Config DecoderIts config.json in plain language Name DecoderWhat 70B, A3B, AWQ or FP8 mean in a model name

Fine-tuneMemory for LoRA, QLoRA and full fine-tuning

What you'll find

Task Fit

Pick a model and a task — chat, code, RAG, agents, reasoning, vision or embeddings — and see, criterion by criterion, how well the model's declared capabilities match the task. It evaluates tuning, modality, size and context, not answer quality.

VRAM Sizing

Model

Task Fit Model Compare Model Anatomy Config Decoder Name Decoder Fine-tune

Serve

Serving Brief vLLM Params Capacity Plan GPU Compare

Learn

LLM 101 Feed VRAM requirements by model Which LLMs fit on which GPU How LLM memory works

How LLM memory works

Three guides to the terms that decide how much GPU memory a language model needs. Each works through real models, and every number in them is computed by the same engine as the calculator.

KV cache size: the formula, with real models

The formula, a worked example, per-token figures for popular models, and why concurrency multiplies it while weight quantization leaves it alone.

GQA, MQA and MLA: how attention decides the KV cache

Four ways to arrange attention heads, why they change the cache by an order of magnitude, and how DeepSeek's latent attention takes the head count out of the formula.

Quantization and VRAM: BF16, FP8, INT8 and INT4

Bytes per parameter from FP32 to INT4, the formats you will meet, why 4-bit is not exactly half a byte, and why it leaves the KV cache alone.