Parameters (AI)
What is Parameters (AI)?
Parameters are the billions of numerical values inside an AI model that encode everything it learned during training. They define the model's capabilities, and adjusting them during training is how the model learns. Users cannot change these directly: they are fixed after training.
At a glance
- Also known as
- Weights (specifically the connection values)Model weightsLearned parameters
- Used for
- Encoding everything a model has learned from training dataDetermining the capability, style, and behaviour of a modelThe fundamental component that defines a trained AI model
- Common tools
- All AI models and neural networksPyTorch and TensorFlow (training and parameter management)Hugging face model hub (parameter storage and sharing)
- Related terms
- Neural networkTrainingFine-tuningModelWeightsCFG scaleInference
- How it works in simple terms
- During training, the model's parameters are adjusted millions of times to minimise errors. After training, they are fixed. When you use a model to generate an image or text, your input passes through the model's architecture, transformed at every layer by these fixed parameter values, until an output emerges.
- Where you encounter this
- Every AI generation tool is built on a model with a specific parameter count. Model descriptions often include parameter counts (e.g. '7B parameters', '70B parameters') as an indicator of scale. In generation interfaces, adjustable settings like CFG scale and steps are sometimes colloquially called 'parameters'.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
How it compares
Compared with related concepts
Model parameters and generation parameters (also called inference parameters or sampling parameters) are often conflated but are technically distinct. Model parameters are the fixed, learned internal values that define what the model can do: they are the model itself, and cannot be changed by users. Generation parameters are user-adjustable settings that control how the model's fixed capabilities are applied to a specific generation request: they shape the output without altering the underlying model. Changing generation parameters changes how the model performs; changing model parameters (through fine-tuning or retraining) changes what the model can do.
Think of it like…
Parameters in an AI model are like the accumulated knowledge in an expert's brain: developed through years of study and experience, encoding everything they know about their field in ways they could not fully articulate explicitly. When they are asked a question, that deep, encoded knowledge shapes their answer. The question itself (the prompt) is the immediate input; the accumulated knowledge (the parameters) is what converts that input into a meaningful response.
Pro tip
When exploring different models for a generation task, parameter count is useful context but should not be the primary selection criterion. A well-trained 7B parameter model often outperforms a poorly trained 70B model on specific tasks. Focus first on the model's demonstrated outputs in your target domain: what it has been trained on, what styles and quality levels it produces: rather than its raw parameter count. Parameter count is an indicator of capacity, not a guarantee of quality.
Types and variations
- Weights are the parameters defining the strength of connections between neurons: the most numerous type of parameter in most networks.
- Biases are additional parameters added at each neuron that shift the activation function independently of the input, providing additional flexibility.
- Hyperparameters are settings that define the training process itself ( learning rate, batch size, number of training epochs ) rather than the model's learned values; they are fixed before training begins, not learned from data.
- Generation parameters (CFG scale, steps, seed, sampler) are user-accessible settings that control how a trained model is applied to a specific generation task, distinct from the model's internal learned parameters.
Ready to make your first scene in Morphic?
Try MorphicCommon use cases
- Understanding model parameters is relevant when comparing models by scale (a 70B parameter model vs a 7B model in the same family), when evaluating fine-tuned models (which begin with a pre-trained model's parameters and adjust them further for a specific domain), when considering the computational requirements of running a model (larger parameter counts require more memory and compute), when interpreting generation quality differences between model versions, and when adjusting generation parameters (CFG scale, steps, etc.
- ) to control how a model's learned parameters are applied to produce a specific output.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
FAQs
Parameters are the internal numerical values ( primarily weights and biases ) that define a trained AI model's behaviour. They encode everything the model learned from training data and determine how the model transforms any given input into an output. After training, parameters are fixed and constitute the saved model. Modern large models contain billions of parameters.
Parameters are the values learned during training: the weights and biases inside the network that are adjusted through the training process to minimise errors. Hyperparameters are settings that define the training process itself ( learning rate, batch size, number of training steps ) and are set before training begins rather than being learned from data. Parameters are outcomes of training; hyperparameters shape how training is conducted.
Not necessarily. More parameters means greater representational capacity: the model can potentially learn more complex and nuanced patterns: but quality also depends heavily on training data quality, architecture design, and training methodology. A smaller model trained on excellent, curated domain-specific data can outperform a much larger general model on specific tasks within that domain. Parameter count is an indicator of scale, not a guarantee of quality or suitability for a specific use case.
Weights are the specific type of parameters that define the strength of connections between neurons in a neural network: the most numerous and significant type of parameter. In common usage, 'model weights' and 'model parameters' are often used interchangeably to refer to the complete set of learned values that define a trained model. When a model is downloaded or shared, the file being transferred contains these weights.
Generation parameters (also called inference or sampling parameters) are user-adjustable settings in a generation interface ( CFG scale, steps, seed, noise level, sampler ) that control how a trained model is applied to a specific generation task. Model parameters are the fixed, learned internal values that define what the model can do. Generation parameters shape how the model's capabilities are expressed; model parameters define what those capabilities are.
Fine-tuning starts with a pre-trained model's existing parameters and continues the training process on a smaller, specialised dataset, adjusting the parameters further to improve performance on that specific domain or task. The result is a fine-tuned model whose parameters encode both the broad capabilities learned during pre-training and the specialised knowledge developed during fine-tuning. Fine-tuning is significantly less computationally expensive than training from scratch because the pre-trained parameters provide a strong starting point.
Each parameter requires memory to store and computation to apply during inference. A model with 70 billion parameters requires approximately 140GB of memory to store in 16-bit precision (each parameter stored as a 2-byte value). Processing an input through a 70B model requires performing billions of mathematical operations, requiring powerful hardware. This is why very large models typically run on cloud infrastructure with specialised hardware rather than on consumer devices.
Not in standard consumer interfaces. In normal use, model parameters are fixed and inaccessible: users interact with the model through prompts and generation parameters without directly touching the underlying weights. Advanced users with access to the model files and appropriate hardware can perform fine-tuning, which adjusts parameters through additional training. Some platforms and techniques like LoRA offer more accessible parameter adjustment methods that modify a small subset of parameters efficiently.
