NAM A2 Architecture
NAM A2 Architecture
A2 is the second-generation standard neural-network architecture for Neural Amp Modeler snapshot models.
A2 was officially released on June 2, 2026 and became NAM's new standard recipe for modeling guitar and bass amplifiers and pedals.
It supersedes the original standard NAM WaveNet family, now collectively referred to as A1, while preserving support for existing A1 models.
A2 was developed with three major goals:
- Maintain or improve modeling accuracy
- Reduce computational requirements
- Provide a more flexible architecture for software and hardware implementations
A1 and A2
Before A2, NAM's commonly used WaveNet model sizes were known as:
- Standard
- Lite
- Feather
- Nano
During development of A2, these earlier architectures were renamed:
- A1-standard
- A1-lite
- A1-feather
- A1-nano
The term A1 therefore refers retrospectively to the original standard NAM architecture family.
A2 is the succeeding generation.
Existing A1 models did not become obsolete when A2 was introduced.
Why A2 was developed
NAM began as an open-source research and software project, but its models eventually appeared across a much wider range of systems.
NAM playback now occurs on:
- Windows computers
- macOS computers
- Linux computers
- Single-board computers
- Embedded processors
- Dedicated guitar hardware
- Web services
- Commercial audio products
These systems have very different computational capabilities.
An architecture appropriate for a desktop computer may consume too much processing power in an embedded device, while an architecture optimized solely for very small hardware might unnecessarily compromise accuracy on more capable systems.
A2 was developed with these different environments in mind.
Industry-informed development
A2 was deliberately developed with input from companies and developers implementing NAM in real products.
During development, participating builders measured candidate models on their target hardware and reported computational performance.
This allowed architecture decisions to be evaluated across different processors rather than optimized solely on the developer's computer.
This was a significant change from the development process of the original A1 architecture.
Computational efficiency
One of the primary objectives of A2 was improved computational efficiency.
The development process evaluated the relationship between:
- Model accuracy
- CPU usage
- Model architecture
- Model size
- Target hardware
The goal was not simply to create the smallest possible neural network.
Instead, the objective was to obtain a better relationship between computational cost and modeling accuracy.
This is especially important for hardware capable of running several effects or models simultaneously.
Modeling accuracy
Reducing CPU usage is useful only if the resulting model remains sufficiently accurate.
A2 development therefore evaluated both computational performance and modeling quality.
NAM's stated objectives for the new generation included maintaining or improving:
- Modeling accuracy
- CPU efficiency
- Training time
The final architecture was selected through technical evaluation followed by subjective listening tests.
Evaluation using musical material
A2 development emphasized evaluation using musical input rather than relying exclusively on synthetic test signals.
Guitar direct recordings and signals processed through common pedals were used as relevant evaluation material.
This reflects an important distinction:
Training data and evaluation data do not necessarily serve the same purpose.
A model may be trained using a carefully designed excitation signal while its practical performance is evaluated using the kind of musical material it will actually process.
Wide range of equipment
A2 was intended to remain a general-purpose default architecture rather than requiring a different neural-network recipe for every amplifier or pedal.
Development therefore considered a variety of guitar and bass equipment.
The primary target remained nonlinear equipment such as:
- Amplifiers
- Preamps
- Overdrive pedals
- Distortion pedals
- Related guitar and bass equipment
Speaker cabinets and microphones present a different modeling problem and can often be represented efficiently using techniques such as impulse responses.
See Cabinets and IRs.
A2 and WaveNet
A2 remains within the broader WaveNet-based architecture implemented by NeuralAmpModelerCore.
Development of A2 required extending the capabilities of the NAM real-time DSP core so that new architecture configurations could be represented efficiently.
The final A2 recipe uses capabilities introduced during this development process rather than merely changing the number of channels in an A1 model.
Consequently, an older NAM player cannot necessarily run A2 simply because it can load A1 WaveNet models.
NeuralAmpModelerCore
NeuralAmpModelerCore is the open-source real-time DSP library used to execute NAM models.
A2 development required new capabilities in the core library.
At release, the NAM project specified the following minimum relevant versions for A2 support:
- neural-amp-modeler v0.13.0
- NeuralAmpModelerCore v0.5.2
- NeuralAmpModelerPlugin v0.7.14
Developers implementing NAM playback should consult the current project documentation because these versions will continue to advance.
A2 is not forward-compatible with old players
An important distinction exists between backward compatibility and forward compatibility.
A current A2-capable NAM player can continue to support older A1 models.
However, a player built before A2 support was implemented cannot be expected to run A2 models.
In other words:
New player → A1 model: supported
New player → A2 model: supported
Old A1-only player → A2 model: not necessarily supported
An A1-capable product generally requires a software or firmware update to add A2 support.
The NAM project stated during A2 development that there is nothing inherent in A2's CPU requirements that prevents existing NAM hardware from supporting it, provided the manufacturer implements the required software changes.
Existing A1 models remain valid
A2 does not invalidate existing NAM collections.
A1 models remain usable in compatible current NAM software and hardware.
This is particularly important because large libraries of A1 captures were created before A2 existed.
The introduction of A2 should therefore be understood as an expansion of the NAM ecosystem rather than a replacement that makes older models unusable.
Training A2 models
At the time of A2's release, A2 training was available through:
- The official NAM Google Colab workflow
- The local NAM GUI trainer
- TONE3000
Current simplified NAM training workflows use A2 as the standard architecture for newly trained snapshot models.
See Training a NAM Model.
Retraining existing captures
A model creator who preserved the original training input and recorded output can potentially train a new A2 model from the same capture data.
The physical amplifier or pedal does not necessarily need to be captured again.
This demonstrates the value of preserving:
- Original NAM training signal
- Original returned recording
- Calibration information
- Equipment settings
- Capture notes
A capture recording is the measurement of the equipment.
The neural-network model is one interpretation of that measurement.
As training technology improves, preserved capture data may therefore remain useful.
A1 and A2 are separate models
Retraining an A1 capture as A2 creates a new model.
It does not transform the existing A1 model internally into A2.
The A1 and A2 files represent independently trained neural networks, even when both were trained from the same recorded capture data.
They may therefore exhibit small differences.
For meaningful comparison, both should be evaluated using the same source material and playback chain.
Slimmable NAM research
A2 development followed research into Slimmable NAM, a method for creating neural amp models whose computational cost can be adjusted at runtime.
The concept addresses a practical problem:
Different playback devices have different amounts of processing power available.
A desktop computer may be able to run a large neural network easily, while an embedded processor may need a smaller computational workload.
Slimmable modeling explores the possibility of allowing one trained model architecture to operate at different computational sizes.
This research helped inform the goals and development direction of A2.
A2 should not, however, simply be described as another name for Slimmable NAM. A2 is the standard architecture resulting from the broader architecture-development project.
Hardware support
Because NAM is open source, A2 support can be implemented by third-party hardware manufacturers.
Support depends on the firmware and NAM implementation in the particular device.
A product advertised as supporting .nam files should therefore not automatically be assumed to support every NAM architecture.
When evaluating hardware, determine whether it supports:
- A1
- A2
- Both
and verify the firmware version where appropriate.
Example: HeadRush
In August 2026, HeadRush released Firmware 5.1 for the Prime, Core, and Flex Prime processors with native playback support for both NAM A2 and A1 models.
This provides an example of the distinction between file-format recognition and architecture support: hardware already capable of running NAM-related workloads still required updated firmware implementing the newer architecture.
Other NAM-compatible products may have different support status.
Always consult current documentation for the specific product.
Identifying an A2 model
The architecture and configuration information required to reconstruct a NAM model is stored in the .nam file.
Compatible software can therefore determine the model architecture from the file itself.
A filename does not need to contain "A2" for the model to be an A2 model, although creators may include architecture information in filenames or descriptions for convenience.
Why architecture matters to users
For many musicians, the internal neural-network architecture is not something that must be understood in detail.
However, architecture matters when:
- A model will not load
- A hardware device supports A1 but not A2
- CPU usage is important
- Models are being trained
- A1 and A2 versions are being compared
- Software or firmware is being updated
- Models are being distributed to other users
When troubleshooting a model that refuses to load, checking architecture compatibility should be one of the first steps.
Why architecture matters to developers
For developers, architecture support determines how the model must be reconstructed and executed.
A NAM host must understand the architecture and configuration contained in the model file and implement the required processing correctly.
A2 development was conducted partly in cooperation with hardware implementers specifically to make the new standard practical across a wide range of devices.
The authoritative implementation remains the open-source NAM code.
A2 development timeline
Major public milestones included:
November 2025
Slimmable NAM research was presented and the concept of a future A2 standard was discussed.
January 17, 2026
The formal Architecture A2 development project was announced.
January 2026
New NeuralAmpModelerCore capabilities and hardware test models were released for developers.
March 2026
Optimization work produced early candidate A2 designs.
April 2026
Development progressed to final listening tests and updated NeuralAmpModelerCore releases.
June 2, 2026
A2 was officially released as NAM's new standard architecture.
Practical guidance
For newly trained NAM snapshot models, A2 is the current standard.
Keep older A1 models. They remain useful and supported.
When distributing A2 models, remember that some older software and hardware may require an update before they can play them.
When a NAM model fails to load, determine:
- Whether the model is A1 or A2.
- Whether the player supports that architecture.
- Whether the software or firmware is current.
- Whether the file itself is valid.
Do not assume that failure to load an A2 model means the .nam file is corrupt.