// Section 1 — The Model · 2 MIN READ

[✓] VERIFIED MANUAL ENTRY — This concept has been rewritten from primary sources and is legally cleared for production.

Parameters

The internal floating-point weights and biases of a neural network, optimized during training, that define the model's behavior.

Parameters (also called weights and biases) are the billions of floating-point numbers stored inside a model. In a Transformer, they represent the strength of connections between attention heads and feed-forward layers.

Everything a model "knows" globally is compiled directly into these numbers (known as parametric knowledge).

  • Frozen at Inference: During standard API usage (inference), these parameters are read-only.
  • Stateless Execution: Nothing you input, correct, or execute inside a chat session will update these parameters. The next API request starts with the exact same numbers.
  • VRAM Impact: The parameter count directly dictates the VRAM memory footprint required to host the model (e.g., a 7-billion parameter model stored in 16-bit precision requires ~14 GB of VRAM to load).

Fine-Tuning vs. In-Context Learning

To update parameters, you must run a training loop (fine-tuning). To update behavior without changing parameters, you must load context.

Parameters (Weights)  ──► Changed via Fine-Tuning (Permanent, high latency, expensive)
Context Window (Input) ──► Changed via Prompting / RAG (Session-only, immediate, cheap)

# AVOID

Do not attempt to explain project-specific specifications to an agent in the hope that it will "remember" them in a new session.

  • Avoid: "I explained the API change to the agent yesterday; why is it still using the old method today?"
  • Write: "The agent's parameters are frozen. We must inject the new API guidelines into the system prompt of every new session."

# USAGE

Developer A: "Can we get the model to memorize our codebase schema by fine-tuning its parameters?" Developer B: "No, that's too slow and expensive. Let's keep its parameters frozen and use RAG to inject the relevant database schemas directly into its context window instead."

// SEE_ALSO

ModelThe compiled, frozen parameters that calculate next-token probability distributions. A model is completely stateless and pure.
TrainingThe computational process of optimizing a model's parameters by exposing it to datasets and adjusting weights via backpropagation.
InferenceThe execution phase where a trained model processes input tokens to generate output tokens.
Parametric KnowledgeThe frozen world facts and coding capability compiled directly into the model's parameters during training, which cannot be modified during inference.

// SOCRATIC_VALIDATION

Interactive Concept Quiz

QUESTION 1 OF 3SCORE: 0/3

[EDIT_THIS_TERM_ON_GITHUB ↗]