What Is Neural Amp Modeler? The Free Amp Tech That Outscored the Paid Ones

If you have been anywhere near a guitar forum in the last two years you have seen the letters NAM, usually attached to someone insisting their free plugin sounds better than your expensive one. They have a point, and the reason is more interesting than the argument.

The short version

Neural Amp Modeler is free, open-source software that trains a small neural network to behave like one particular real amplifier. Steve Atkinson released it in 2022. You feed it a capture file, it becomes that amp, and you play through it.

Nobody sat down and modelled the circuit. The training process plays a known test signal through the real amp, records what comes back, and then adjusts a network until its output matches the recording closely enough that your ears cannot separate them. The result is a file of a few hundred kilobytes that contains, in a real sense, that amp on that day at those settings.

A capture is not an impulse response

People mix these up constantly, and the distinction explains most of what a rig sounds like.

An impulse response captures something linear, which in practice means a speaker cabinet, a microphone and the room around them. Feed it a signal, it colors it, and the coloring is the same at every volume. An IR cannot distort, because distortion is by definition the part that changes with how hard you hit it.

A NAM capture handles exactly that changing part: the compression as a power section sags, the way breakup arrives earlier when you dig in, the bloom on a held note. Most full rig tones are a capture of the amp followed by an IR of the cab, which is why the libraries are full of both.

What a capture cannot do

You cannot turn the gain knob. A capture is a photograph of one amp at one setting, so the tone stack is baked into the file. That is why the library counts are in the hundreds of thousands: people capture the same head at four gain positions and upload all four, and the ones who do it well label them properly.

You are not stuck, though. EQ before the capture behaves a lot like a pedal in front of an amp, EQ after it behaves like a different mic position, and swapping the cab IR changes the sound more than most people expect. Anyone who has spent an evening on cab choices already knows this.

What changed with A2

A2 is the second NAM architecture, released June 2, 2026, built by TONE3000 with Steve Atkinson. It is a rewrite rather than a tune-up, and it went two directions at once: closer to the source and cheaper to run.

A2-Full sounds better than the original architecture while using 30 to 40 percent less CPU. A2-Lite is the version built for hardware, and it runs at half the CPU of a three-dollar ARM Cortex-M7 chip, which is the detail that matters for where this goes next. An M-series MacBook can hold 64 A2-Full models at once, or 200 Lite ones, which is more amps than anyone needs and a good indication of how little each one costs you.

Because the whole thing is MIT licensed, hardware companies can ship it without paying anyone. Blackstar, Darkglass, HeadRush, Lava Music, Chaos Audio and Dimehead have all added A2 support. That is the real story of the last year: the free format is turning up inside gear you buy.

Where to actually run it

The easiest route is the TONE3000 plugin, released September 2026, which is free, runs standalone with no DAW, and browses the whole capture library from inside the plugin. Steve Atkinson's original NAM plugin is still there and still free if you prefer something plainer that you point at your own files. A pile of other hosts load .nam files too, and Slopsmith built one in.

You will need an audio interface with an instrument input. Not an expensive one: the cheap interface guide covers what actually matters, which is a Hi-Z input and a driver that does not add latency.

Does the free one really beat the paid ones?

On the numbers TONE3000 published, yes. They tested 39 tones on ESR, MAE, LOG_MEL and MRSTFT, then ran a blind listening test where more than a thousand people submitted over a hundred thousand ratings, and A2 finished above Neural DSP Neural Capture V2, IK Multimedia ToneX V2 and Line 6 Proxy on both.

The obvious caveat is that TONE3000 built A2 and ran the test on it. They released the code and the raw data, which makes it checkable rather than a press release, and independent listeners have broadly agreed since. The comparison piece goes through what the test did and did not measure.

Then the boring part

None of this makes anyone play better. A perfect capture of a 1959 Plexi will reproduce a badly timed bend in exquisite detail. I say that as someone who has spent whole evenings auditioning captures and finished with nothing new under my fingers.

The fix is to have something checking the notes while you enjoy the tone. Guitar Quest listens through the same interface your amp sim is running on, grades what you played note by note, and tells you what to work on next, so the rig becomes the reward for practicing rather than a substitute for it. One payment of $30, and the free demo asks for no signup. This is what the drilling looks like:

Related reading

• The TONE3000 plugin is free, and it is very good
• NAM vs ToneX vs Neural Capture
• Free guitar amp sims worth installing
• The best amp modeler and audio interface