Neural Amp Modeler
Neural Amp Modeler
Neural Amp Modeler (NAM) is a free and open-source system for creating highly accurate digital models of nonlinear audio equipment using deep learning.
NAM is commonly used to model guitar and bass amplifiers, pedals, preamps, and related signal chains. Unlike a conventional algorithmic model designed specifically around a particular amplifier or circuit, a NAM model is trained from measured input and output audio representing the behavior of the equipment being modeled.
NAM is both a technology and an open-source ecosystem. Models can be created independently, exchanged as files, and played by software and hardware that implement compatible NAM playback.
Overview
A typical NAM modeling workflow consists of:
- Obtaining the standardized NAM training signal.
- Sending that signal through the equipment to be modeled.
- Recording the equipment's output.
- Training a neural-network model using the original input and recorded output.
- Exporting the trained model as a .nam file.
- Loading that model into compatible playback software or hardware.
This separation between capture, training, model file, and playback is an important characteristic of the NAM ecosystem.
A model does not have to be trained by the device that eventually plays it.
Open-source architecture
NAM is an open-source project.
The project separates several major functions into different software components, including:
- Training code used to create models
- Core DSP used to run models in real time
- Plugin and standalone software used for playback
This architecture allows third-party developers to incorporate NAM playback into their own software and hardware.
It also allows researchers and developers to examine, modify, and extend the underlying technology rather than treating the modeling system as a closed platform.
What NAM can model
NAM is designed for modeling nonlinear audio systems.
Common targets include:
- Guitar amplifiers
- Bass amplifiers
- Overdrive pedals
- Distortion pedals
- Boosts
- Preamps
- Combinations of nonlinear devices
- Larger signal chains
The practical suitability of a particular device depends on whether its behavior can be represented effectively by the model architecture and training process.
NAM should not be assumed to reproduce every possible type of audio processor equally well.
Time-varying effects, long delays, modulation, reverberation, and other systems with behavior fundamentally different from the equipment NAM was designed around may require different modeling approaches.
Snapshot models
Most NAM models used for amplifiers and pedals are snapshot models.
A snapshot model represents the behavior of equipment at the configuration used when the training data was recorded.
For example, if an amplifier is captured with a particular combination of:
- Gain
- Bass
- Middle
- Treble
- Presence
- Channel
- Boost or voicing switches
the resulting NAM represents that measured configuration.
Playback controls can alter the signal around or after the model, but they should not automatically be interpreted as equivalent to changing the corresponding controls on the original amplifier.
For substantially different amplifier settings, separate NAM models are commonly created.
Training data
NAM training requires paired input and output data.
The input is the known signal sent into the equipment.
The output is the response recorded after that signal passes through the equipment being modeled.
The trainer uses the relationship between these signals to learn a mathematical model that predicts how the equipment transforms incoming audio.
For the simplified NAM training workflow, the project provides a standardized training signal.
A typical signal path is:
NAM training signal → audio interface → reamp/capture path → device under test → audio interface → recorded output
See:
Local training
NAM provides tools for training models locally on a computer.
The simplified local GUI trainer allows the user to select the original NAM input file and the recorded output file and then train a model.
The current documented simplified workflow uses 48 kHz, 24-bit WAV data and requires the returned recording to match the length of the supplied input file.
The trainer also attempts to determine and compensate for round-trip latency between the input and output recordings.
More advanced command-line training tools provide greater control over the model architecture, training data, and learning configuration.
See Training a NAM Model.
Cloud and third-party training
NAM models do not have to be trained on the user's own computer.
The NAM project provides a Google Colab workflow for cloud-based training.
Third-party services can also train compatible NAM models. This makes it possible to create NAM captures without maintaining a local machine-learning environment.
The current official NAM site identifies TONE3000 as a community model library and a service capable of training NAM models.
The .nam model file
Trained NAM models are normally distributed as files using the:
.nam
file extension.
A .nam file contains the information required by compatible playback software to reconstruct and run the trained neural-network model.
The format includes information describing the model architecture, configuration, weights, sample rate, and optional metadata.
Model metadata
The NAM file specification supports metadata that can provide useful information about a model and its creation.
Depending on the file and trainer used, metadata may include information such as:
- Model name
- Model creator
- Equipment manufacturer
- Equipment model
- Equipment type
- Tone type
- Training information
- Input calibration
- Output calibration
Metadata is especially valuable when models are distributed publicly because the filename alone may provide little information about how the model was created.
Calibration metadata
NAM supports calibration metadata describing the relationship between digital signal level and the analog levels used during capture.
Two important values are:
- input level — the analog level corresponding to digital full scale at the input of the modeled equipment
- output level — the analog return level corresponding to digital full scale in the recording system
Correct calibration can help compatible playback systems reproduce the gain relationship of the original analog equipment.
Calibration metadata is optional. A model without it can still operate, but the relationship between digital playback level and the original equipment may not be accurately reproduced.
See Gain Staging and Calibration.
A1 architecture
The earlier standard NAM WaveNet architectures are now collectively referred to as A1.
These include architectures historically described as:
- Standard
- Lite
- Feather
- Nano
The A1 terminology was introduced to distinguish the original architecture family from the newer A2 generation.
Existing A1 models remain part of the NAM ecosystem and continue to be supported by compatible NAM implementations.
A2 architecture
A2 is the current standard NAM architecture generation.
A2 was officially released on June 2, 2026.
It superseded A1 as the default architecture for newly trained snapshot models while retaining compatibility with existing A1 models.
Current NAM simplified training workflows, including the official Colab and local GUI trainers, can create A2 models.
A2 support requires sufficiently recent NAM playback implementations.
See NAM A2 Architecture.
Playback
A .nam file does not produce audio by itself.
It must be loaded into software or hardware capable of running the model.
NAM playback can exist in:
- Audio plugins
- Standalone applications
- Hardware processors
- Amplifiers
- Embedded products
- Third-party software
Because NAM's real-time DSP implementation is available as open-source code, manufacturers and developers can integrate NAM playback into their own products.
Not every NAM player necessarily supports every model architecture. Newer architectures such as A2 may require updated playback software or firmware.
See NAM Playback.
Cabinets and impulse responses
A NAM model may or may not include a speaker cabinet and microphone.
A direct amplifier NAM can represent the amplifier without a physical cabinet response. In that case, playback normally requires:
- A cabinet impulse response
- A cabinet model
- A physical guitar cabinet
A NAM captured through a physical cabinet and microphone may already contain those characteristics.
Adding another conventional cabinet IR to such a model may result in unintended double cabinet processing.
See:
Model sharing
Because NAM models are files rather than objects locked to a particular capture device, they can be exchanged between users and compatible playback systems.
This has resulted in large community collections of models representing amplifiers, pedals, and other equipment.
The official NAM website currently directs users to TONE3000 for community model sharing and discovery.
When distributing a model, useful documentation includes:
- Equipment modeled
- Equipment settings
- Whether a cabinet is included
- Capture signal chain
- Calibration information
- Model architecture
- Creator
- Relevant notes about the capture
A large model library becomes considerably more useful when its contents are documented well.
NAM as an ecosystem
NAM is unusual among capture technologies because no single manufacturer controls every stage of the process.
The ecosystem can include separate developers and services for:
- Capture
- Training
- Model hosting
- Playback software
- Playback hardware
- Model management
- Research
A model created using one NAM-compatible workflow can potentially be used by many different NAM-compatible playback systems.
This interoperability is one of the most significant characteristics of the platform.
Limitations
A successful NAM training run does not guarantee a perceptually perfect model.
Accuracy can be affected by:
- Training data
- Calibration
- Noise
- Clipping
- Latency alignment
- Model architecture
- Equipment behavior
- Capture signal chain
- Training configuration
Some equipment may also exhibit behavior that is more difficult to reproduce than others.
The resulting model should therefore be evaluated against the original equipment rather than judged solely by whether training completed successfully.