Loading a NAM model
How to import a Neural Amp Modeler capture into Axion, how it boots in the browser, and what to do when a file won't load.
A NAM model is a captured snapshot of a real amplifier — input gain stage, EQ, output stage, all of it — produced by training a neural network on hours of recorded test signal through the actual hardware. The .nam file is the trained weights. Axion runs that network in-browser via a Rust + WebAssembly implementation, so loading a NAM capture means you're playing through that amp, with no plugin, no driver, and no desktop install.
NAM loading is free — every tier can import and play captures. The analytical amp voicings sit alongside it as the always-available default. (Axion is built in step with the NAM community, which is itself free and open.)
Where NAM captures come from
The NAM ecosystem is community-driven. Capture artists run real amps through the NAM Trainer (desktop, free) and publish the resulting .nam files. Public repositories like Tone3000 host thousands of captures; many capture artists also sell commercial packs of their own captures.
Axion plays any .nam file that the open NAM format defines — the LSTM and WaveNet variants are both supported, including the current-generation NAM A2 architecture. Captures produced by the latest NAM Trainer load without conversion.
A2 models ship in two sizes — A2-Full (highest fidelity) and the lighter A2-Lite — and load exactly like any other capture; A2-Lite is the one to reach for on the mobile / lower-headroom path since it's roughly half the CPU.
NOTE Axion never transmits or stores your NAM files on our servers. They live in your browser's IndexedDB, and the only thing that ever rides over the network is a short hash reference if you share a preset that uses one. Your captures are yours.
Importing a capture
Two ways in:
- Drag and drop. Drop a
.namfile directly onto the amp section. The model imports and immediately becomes the active capture. - From the NAM library. Open the amp's source toggle and choose Manage NAM library. The panel that opens lists every imported capture, with an Import button at the top that opens the file picker.
On first load Axion compiles the model's forward pass into a stack-allocated WASM module and caches the result in IndexedDB. The compilation takes a few hundred milliseconds the first time. On every subsequent load the cached module hydrates in under ~80 ms.
Switching, A/B-ing, and removing captures
The NAM library panel shows every imported capture with its size, hash, and import date. Tap a row to make it the active capture; tap Remove to delete it.
The amp's source toggle (top of the amp section) flips between the analytical voicing and the current NAM. This is a real switch — you can A/B mid-strum to compare. Both halves of the source state persist with the preset, so a preset can store "NAM loaded but switched to analytical" if that's what you want.
Removing a capture that's referenced by a saved preset prompts before deleting. If you delete anyway, presets that referenced it fall back to the analytical voicing on load with a small "missing NAM" badge next to the amp name.
When a file won't load
A handful of legitimate causes, ranked by frequency:
- Old NAM format. The format has gone through revisions; pre-0.7 captures may not load. The fix is to re-export the file from the latest NAM Trainer (the original training data is what matters, not the file).
- Truncated file. A
.namthat downloaded partially (interrupted connection) fails the format-validation step. Re-download from the source. - Wrong file type. Some sites distribute
.namfiles inside a.zipor.rararchive. Unpack first; Axion only reads the raw.nam. - Browser storage cap hit. IndexedDB has per-origin quotas (typically several hundred MB to a few GB depending on browser and disk). If you've imported dozens of large captures, the next import may fail to cache. Free space from the library panel.
When the audio drops out under a NAM
NAM inference is the most CPU-expensive thing in the rig — significantly more than the analytical amps or the cab convolution. If you hear crackle or dropouts that only happen with a NAM loaded, the model is the culprit.
Capture artists usually publish two flavours of each capture: the full model (best fidelity, highest CPU) and a feather or lite variant (slightly less detail, much cheaper). Try the feather version first on lower-end machines. On the Xiaomi mobile baseline, feather captures play; full captures usually don't.
A larger buffer also helps — see Audio troubleshooting for the buffer adjustment flow.
NAMs in shared presets
When you publish a preset that uses a NAM (via Sharing a preset), the NAM binary itself is never uploaded — only its content hash. Recipients hear the NAM only if they already have that exact capture imported; otherwise they hear the analytical fallback baked into the preset, with an inline prompt to import the original capture. Loading NAM captures is free on every tier, so anyone who has the file can play it.
This keeps Axion out of the business of redistributing captures, which is important — many community captures are licensed CC-BY-SA, and commercial packs come with their own terms. The hash reference doesn't carry the model; only someone who legitimately owns the capture file can play it.