Neural Amp Modeler is an open-source project that uses deep learning to make models of guitar amplifiers and pedals. It is free, it is MIT licensed, and the core repository has 3,116 stars — created in January 2019 and still being pushed to this year, with the trainer at v0.13.0 as of June 2026.
It is also now a hardware feature. On 26 October 2026 Zoom announced the Neurig, a €349 floor modeler with two touchscreens, battery power and native NAM A2 Lite support. That is the context for learning this now: captures you make are starting to run on pedals, not just plugins.
The one thing to understand first
A NAM capture is a model of your amp at the settings it was recorded at.
Not a model of your amp. A model of your amp with the gain at 6, the bass at 5, the mids at 7, the treble at 4, that speaker, that microphone, in that position, in that room.
Turn the virtual knobs afterwards and you are not getting what the real amp would do at those settings — because the model never saw those settings. Gearnews put the distinction well in its coverage of the Zoom announcement: a NAM capture is “an extremely accurate sonic photograph, but when you change the knobs, the results will not be the same.”
This is why the Neurig ships NAM support and Zoom’s own DNA (Dynamic Neural Archive) engine alongside it — 30 DNA amp models plus drive pedals, 22 conventional models, 22 multi-layer IRs and 139 effects. DNA aims to model an amp so it responds correctly when you turn knobs, which is what Neural DSP and Line 6 do and what NAM, by design, does not attempt.
So the mental model is: NAM is a photograph, not a simulation. Photographs are extraordinarily accurate about the moment they captured and tell you nothing about any other moment. Once that lands, everything else about NAM makes sense — including why people publish dozens of captures of the same amp.
What you need
To play captures, nothing but a computer: the plugin plus someone else’s capture file.
To make captures:
- an audio interface with a reamp-capable output and an instrument input
- a way to get a line-level signal into your amp — a reamp box, or an interface output you can attenuate
- a load box or attenuator, if you are capturing a valve amp through its speaker output, because running a valve power amp without a load will damage it
- a microphone on the cabinet, or a line out / DI from the amp
- the NAM trainer and the NAM plugin
That load box warning is the only genuinely expensive mistake available here, so it goes before anything else.
How a capture works
The method is the same idea as measuring any black box: send a known signal through it and compare what comes out.
- You play a standardised input file through your amp. It contains sweeps, tones and musical material chosen to exercise the amp across frequency and level.
- You record the output.
- The trainer takes the input/output pair and fits a neural network that maps one to the other.
- The result is a
.namfile you load into the plugin.
Because the model is learned from the pair rather than derived from circuit theory, it captures whatever the chain did — including the cabinet, the microphone, the mic position, and the room, if you recorded in one.
That is a feature and a trap. Everything in the signal path becomes part of the capture, permanently. If you want to swap cabinets later, capture the amp through a load box with no speaker and use a separate impulse response for the cab.
Doing it
1. Get the input file. From the NAM project. Do not normalise it, trim it, or re-encode it. The trainer aligns input against output and expects the file it shipped.
2. Set levels and leave them. Set your amp where you want it, set your interface gain so the recorded output peaks comfortably below clipping, and then do not touch anything until the recording is finished. A gain change mid-capture produces a model of two different amps averaged together.
3. Play the file through the amp and record. One pass, start to finish.
4. Check the recording before training. Listen to it. Look at the waveform. Clipping, dropouts, a hum that started halfway through, or a phone buzzing the pickup — all of it gets learned faithfully. Training takes a while; listening takes a minute.
5. Train. The trainer reports a loss figure as it goes. Lower is closer. Treat it as a smoke test rather than a score — a low loss on a bad recording is a precise model of a bad recording.
6. A/B it honestly. Load the capture, play through it, and compare against the real amp at the same settings with levels matched. Matching levels is not optional; louder reliably sounds better, and an unmatched comparison tells you nothing.
Sweep method or dry/wet method
You will see both discussed. The practical difference is how the timing alignment between input and output is established — one relies on the structure of the test signal, the other on recording a dry reference alongside the amp output. The dry/wet approach is more forgiving when latency through your chain is unknown or variable, which is the usual case with digital reamping. Both are documented by TONE3000 with video walkthroughs; pick one and follow it exactly rather than mixing steps.
NAM or an impulse response?
They are not competitors, and knowing which you need saves a lot of effort.
| Impulse response | NAM capture | |
|---|---|---|
| Models | Linear systems: cabinets, mics, reverbs | Non-linear systems: preamps, power amps, drive pedals |
| Captures distortion | No | Yes |
| File size | Kilobytes | A few megabytes |
| CPU cost | Very low | Noticeable |
A speaker cabinet is close enough to linear that an IR describes it well, which is why IRs have worked for cabs for twenty years. An overdriven valve amp is defined by its non-linearity, which is precisely what an IR cannot represent and what NAM exists for.
The common production setup is both: a NAM capture of the amp taken through a load box, plus an IR for the cabinet. That keeps the cab swappable, keeps CPU down, and models the part that actually needs a neural network.
Where the captures are
TONE3000 hosts the community capture library, and it is the reason this ecosystem matters — thousands of amps, including ones you will never have access to, captured by people who own them. The Zoom Neurig reaches it through Zoom’s free SessionZ Pedal app for Android and iOS, which is a notable thing in itself: a hardware manufacturer shipping a route into a community-run archive rather than only its own store.
Before you spend a weekend capturing, search there. Someone has probably already captured your amp, and if the thing you want is a specific knob setting of a common amp, the fastest path is to find a capture near it.
A licensing footnote
Both core repositories — the trainer sdatkinson/neural-amp-modeler and sdatkinson/NeuralAmpModelerPlugin — are MIT licensed, which is why commercial hardware can ship native support without negotiating anything. The project’s own site has a “Build: Use NAM in your product” section, so this is intended rather than tolerated.
What is not settled by that licence is the status of captures themselves. A capture is a derived representation of a specific commercial amplifier, and whether publishing one implicates the amp manufacturer’s trademarks or rights is an open question that the MIT licence on the software does not answer. In practice the community publishes captures freely and names the amps; if you are shipping a commercial product built on captures of a named amp, that is a question for a lawyer rather than a forum.
Sources
- Neural Amp Modeler
- sdatkinson/neural-amp-modeler — GitHub
- How to Capture Your Guitar/Bass Tones For Free with Neural Amp Modeler and TONE3000 — TONE3000
- TONE3000
- Zoom Neurig: Compact Touchscreen Modeler With Native NAM A2 and Battery Power — Gearnews