NAM is coming to Fractal

Neural Amp Modeler (NAM) “is an open-source project that uses deep learning to create models of guitar amplifiers and pedals with state-of-the-art accuracy,” started by Steven Atkinson. It uses machine learning to make a digital “capture¹” of analog gear such as a guitar or bass amp, a pedal, a piece of outboard equipment, or a whole signal chain. A NAM model is a snapshot of that gear at one set of knob positions. In contrast with earlier profilers like Kemper and ToneX that kept their formats closed, NAM’s trainer, core DSP library, and plugin are all open. That has let hardware makers build NAM playback into their own products, and it produced a very large community library of captures, mostly hosted on TONE3000. 

In June 2026, TONE3000 and Atkinson released Architecture 2 (A2), the first new default network since the original A1. It comes in two versions. A2-Full is meant for DAW and studio use, and TONE3000 says it uses 30 to 40 percent less CPU than A1 while sounding more accurate. A2-Lite is built for embedded hardware and reportedly runs at about 50 percent load on a roughly $3 (bulk) ARM Cortex-M7 600 MHz chip. Capture-quality modeling can now run on inexpensive pedals and amps, which has started what appears to be a sea change in the industry. Dozens of products/vendors support A2, especially A2-Lite, already, and more are being added almost daily. Even my year-old, definition of inexpensive, NUX MG-300 Mk II got a firmware update several months ago that enabled A2-Lite. So do the MG-30 and MG-400.

While other higher end products, like Line 6 (Proxy), Neural DSP’s Cortex products (Neural Capture V2), etc., don’t support A2 (yet?), gobs of lower end devices already do, including, e.g., SKUs from HeadRush and Darkglass (A2-Full), and, e.g., the Hotone Ampero II, Mooer GE/GS1000, Valeton GP-150/180, Blackstar, and VTR (A2-Lite).

I don’t have any immediate plans to use A2 in my Fractal FM3. Heck, I don”t even know if it will be made available (“too early to say”). But Fractal founder Cliff Chase teased it running on the AxeFX III on September 17th, 2026, so it’s at least possible it will trickle down to their lower spec units. Interesting to see that develop. So far, the only “legacy limitation” I’ve seen is that the older Mk I will lose any IRs stored in USER 2 bank (which is being deprecated) when loading .nam files (which are JSON; Fractal’s software converts them into binary sysex files).

Teaser photo showing an AxeFX with a NAM block

Cool new tool added to the toolbox, potentially, eventually.

_____

¹ The capture process is simple in concept. You download a standard sweep/test signal from TONE3000 and send it from your audio interface through a reamp box into the gear. Then you record the output with peaks around -8 dB and export it as a 24-bit, 48 kHz mono WAV that is exactly as long as the original signal. Next you upload that file to TONE3000's cloud trainer (TONEZONE3000), which runs on NVIDIA RTX 4090 GPUs and is free. The trainer learns how the recorded output relates to the clean input, by default over 400 epochs. It then reports an error-to-signal ratio (ESR), where lower is better and anything under about 0.01 counts as excellent. The finished model can be downloaded or published to the library.

Captures are hit or miss, though, mainly because anyone can make and upload one, and the quality depends on the person capturing. Clipped recording levels, ground-loop hum or other noise, a poor reamp setup, and mismatched gain staging all get "learned" into the model as if they were part of the amp's sound. A good ESR only tells you the model matches that particular recording. It doesn't mean the recording was a good one or that the amp sounded good at those settings. Each capture is also frozen at one set of knob positions, so a model captured with the gain up won't behave like the amp with the gain turned down. Uploaders are also inconsistent about whether a cabinet or mic is included, so similar-sounding names can produce very different results. On top of that, the A2 upgrade has created two tiers of models: retrained ones that really improved and converted ones that didn't. You usually need to audition several captures of the same amp before you find a good one.