What Is Neural Amp Modeler (NAM)?

Neural Amp Modeler — usually shortened to NAM — is an open-source neural-network format for capturing the sound of a real guitar amplifier and reproducing it in real time. This page explains what NAM actually is, how the underlying .nam files work, and how it's different from the amp sims and profiling gear you may already know.

The short version

NAM guitar tone is captured, not designed. Instead of an engineer hand-tuning a circuit model to approximate an amplifier, a neural network is trained directly on recordings of the real amp — it learns the amp's exact input/output behavior across thousands of signal levels and frequencies. The result is a small model file (a .nam file) that reproduces the amp's tone with remarkable fidelity, and can run in real time on a phone.

How a .nam file is made

Capturing an amp for NAM follows a repeatable process:

1. A wide-bandwidth test signal (a long, carefully designed audio sweep) is played into the real amplifier through an audio interface.
2. The amp's output — through a mic'd cabinet, a load box, or direct from a modeler — is recorded back in.
3. A neural network is trained on the input/output pair until it can predict the amp's output from any input signal, not just the training sweep.
4. The trained network's weights are exported as a .nam file — typically a few hundred kilobytes to a few megabytes, depending on the architecture.

Anyone can run this capture process with free, open-source tooling, which is why thousands of community-captured .nam models exist for everything from boutique tube heads to vintage combo amps — many of them shared on platforms like Tone3000.

WaveNet and LSTM: the two common architectures

Most .nam files use one of two neural network architectures:

• WaveNet — a dilated convolutional network (originally developed for speech synthesis) adapted to model audio waveforms sample-by-sample. WaveNet captures tend to reproduce nonlinear distortion and harmonic content with high accuracy.
• LSTM (Long Short-Term Memory) — a recurrent neural network architecture well-suited to modeling how an amplifier's response depends on recent signal history (important for dynamics and sag).

Ampz's inference engine loads either architecture — plus the newer slimmable "A2" container format, which lets a single model run at a lower internal resolution for less CPU cost — with no conversion step and no loss of fidelity versus the original capture.

How NAM differs from a traditional amp sim

Traditional software amp simulators (the kind that have existed since the early 2000s) work by modeling the amplifier's electronic circuit — the preamp tubes, the power section, the tone stack — as a chain of discrete DSP components tuned by ear to approximate the target amp. This works well, but it's fundamentally an approximation built by a developer's judgment about what "sounds right."

NAM instead learns the amp's actual measured behavior. There's no circuit model in the loop — the neural network's weights are the tone. This tends to produce a more faithful match to the specific unit that was captured, including its specific quirks, at the cost of not being adjustable in the way a circuit model's virtual components are (you can't turn a "virtual power tube bias" knob on a NAM capture, because there isn't one — the amp's front-panel knob settings are baked into whichever capture you loaded).

How NAM differs from a hardware profiler (Kemper, ToneX)

Hardware profilers like the Kemper Profiler or IK Multimedia's ToneX pedal use a broadly similar idea — measuring a real amp and reproducing its behavior — but they're closed, proprietary systems tied to specific hardware or apps. NAM is an open file format and an open-source inference library: any developer can build a NAM player, any amp owner can capture their own amp with free tools, and the resulting .nam files are portable between every app that supports them.

The other key difference is where the model runs. Profiler captures typically live on dedicated hardware. NAM was built to run efficiently enough for real-time inference on general-purpose hardware — including, with the right native inference engine, an iPhone. Ampz's C++ engine is built specifically to hit sub-2.5 ms round-trip latency running .nam files directly on iOS, with no cloud processing and no external hardware required beyond an audio interface.

Where to get .nam files

You don't need to capture your own amp to use NAM. Tone3000 is a large community library of free, community-captured .nam models and cabinet impulse responses (IRs) covering everything from clean Fender-style combos to high-gain metal heads. Ampz has a built-in Tone3000 browser, so you can search, preview, and download directly inside the app.