← HomeGUIDE 02 / GGUF

GGUF beginner guide: models, quantization, RAM, and VRAM

Learn what a GGUF file is, how quantization names work, how much memory local AI needs, and how to choose a practical model for Local Idea Studio.

What GGUF means

GGUF is a portable file format used by llama.cpp-compatible runtimes. One file can contain model weights and metadata needed for local inference. Local Idea Studio runs compatible GGUF files through node-llama-cpp, which is built on llama.cpp.

Model size and parameter count

Names such as 1B, 3B, 7B, and 14B roughly describe the number of model parameters. More parameters can improve capability, but also increase download size, memory use, and generation time. A well-chosen small model is often more useful than a large model that barely fits.

Quantization in plain language

Quantization compresses model weights so they need less storage and memory. The tradeoff is some loss of precision. Common GGUF filenames include labels like Q2_K, Q4_K_M, Q5_K_M, Q6_K, and Q8_0.

QuantizationTypical tradeoffGood use
Q2 / Q3Smallest files, larger quality reductionVery limited memory or quick experiments
Q4_K_MStrong balance of quality and sizeRecommended starting point for many users
Q5_K_MMore memory, usually better fidelityWhen the model fits comfortably
Q6 / Q8Large files, closer to higher precisionSystems with plenty of RAM or VRAM

How much memory is needed?

The GGUF file size is a useful starting estimate, not the total requirement. The runtime also needs memory for context, temporary work buffers, and the application. GPU offload needs VRAM; CPU inference primarily uses system RAM.

  • Keep several gigabytes free beyond the model file size.
  • Longer context consumes more memory.
  • Flash attention can reduce memory use on supported backends.
  • If a model fails to load, lower context, use a smaller quantization, or reduce GPU layers.

Context is not output length

Context is the total working window for system instructions, conversation history, prompt, and generated reply. Maximum output tokens only limits the new response. Local Idea Studio caps context to the model’s reported trained limit and keeps per-model settings so a small model does not inherit an unsafe value from a larger one.

Choosing your first model

  1. Pick an instruction or chat-tuned model, not a raw base model.
  2. Start with a size appropriate for your memory.
  3. Choose Q4_K_M or Q5_K_M when available.
  4. Read the model card and license on Hugging Face.
  5. Start at a modest 4K context and increase only when needed.
A GGUF extension alone does not guarantee compatibility. The model architecture must also be supported by the bundled llama.cpp runtime.