700,000 Amp Tones Just Went Free. The Plugin Is Version 0.0.4.

TL;DR

On September 1, TONE3000 released a free, MIT-licensed plugin and standalone app that browses and loads its library of more than 700,000 neural amp captures and impulse responses directly inside your DAW. The models come from NAM A2, an open architecture the company built with Neural Amp Modeler creator Steve Atkinson, which is small enough to run at 50% CPU on a $3 ARM Cortex-M7. The commercial amp-modeling business has spent a decade selling captures of the same gear. The capture layer, the model layer, and now the host layer are all free and open.

What actually shipped

The TONE3000 Plugin is a JUCE-based C++ host that ships as VST3, AU, AAX, CLAP and LV2, plus a standalone app, on macOS (Apple Silicon and Intel), Windows and Linux. That is essentially every plugin format that exists, which is a slightly unusual amount of build matrix for a first public release.

Inside, you get unlimited stacking of NAM captures and IRs, dual chains with branching and panning, stereo processing with auto-align and auto-balance, pre-block, post-block and global parametric EQ, oversampling, calibration, MIDI mapping, and a tuner. You log in with a TONE3000 account and the library streams in; you can also drop in local .nam files.

The version number is 0.0.4. The company says the library behind it serves more than 600,000 monthly active users and has passed nine million downloads. Shipping a 0.0.4 to a six-figure user base is a choice, and the CEO more or less says so out loud.

"Because the TONE3000 Plugin is open source, right now is the worst it will ever be." - Stanley Vergilis, co-founder and CEO

The repository is C++ with a TypeScript and React UI layer, MIT-licensed, and was created in September 2025, so this is a public unveiling of something that has been cooking for about a year. It pulls Steve Atkinson's original NAM plugin lineage forward: NeuralAmpModelerCore and AudioDSPTools sit in-tree.

The model underneath

The interesting part is not the host, it is NAM A2, launched June 2, 2026. A neural amp capture is a small network trained to reproduce what a specific piece of analog gear does to a signal: you play a known reference tone through the real amp, record the output, and train the model to match the difference. The result is a file, not a preset. Think of it less as sampling a note and more as photographing the amp's reflexes, which is why it holds up when you change your playing instead of falling apart the way a static EQ curve would.

A2 is a 23-layer network across three dilation stacks, with a receptive field of roughly 6,350 samples, about 132 ms at its native 48 kHz. It ships in two sizes: A2-Full uses an 8-channel network, A2-Lite uses 3. Both are a single file. Oversampling at 2x, 4x and 8x is available with minimum-phase half-band filters for players running at other rates.

A2-Full is claimed to be 30-40% cheaper on CPU than the previous A1-Standard, meaning you can run three A2-Full instances for the cost of two old ones. On a MacBook M-series the company quotes 64 simultaneous A2-Full models, or 200 A2-Lite. A2-Lite hits a 316x realtime factor on an M5 with its optimized fast path.

The blind test, and how much to trust it

TONE3000 ran a MUSHRA listening study in March 2026, the same subjective-quality methodology the BBC and EBU use for codec evaluation. More than 1,000 participants produced over 100,000 ratings, comparing recordings of real gear against anonymized digital models. Objective scoring ran in parallel across 39 tones, covering guitar and bass amps, pedals, outboard gear and full chains, using ESR, MAE, LOG_MEL and MRSTFT with Bayesian Elo aggregation.

The competitors were Neural DSP Neural Capture V2, IK Multimedia ToneX V2, and Line 6 Proxy. TONE3000 reports A2 on top by a wide margin in both the objective sweep and the listening test.

Here is the caveat, stated plainly: TONE3000 designed the test, trained the models, and published the results. The methodology is legitimate and the participant count is unusually large for this corner of the industry, but the exact median scores are published as a chart rather than as a table of numbers, and no independent lab has replicated it. Treat "beats Neural DSP" as a vendor claim with an unusually good paper trail behind it, not as settled fact. The upside of the model being open is that anyone who disagrees can go train their own and post the file.

Why the $3 chip matters more than the plugin

A free plugin is nice. A free plugin whose model format also runs on a 600 MHz Cortex-M7 that costs three dollars in bulk is a different kind of event, because it means the same .nam file a producer loads in Logic can ship inside a pedal. The compact .namb binary format exists specifically for that embedded path.

That path is not hypothetical. NAM already runs on shipping hardware from HeadRush (Prime, Core, Flex Prime), Blackstar (Beam Mini, Beam Solo), Darkglass (Anagram), LAVA MUSIC (LAVA STUDIO) and Chaos Audio (Stratus, Nimbus). The capture format is quietly becoming the JPEG of guitar tone: not the best conceivable container, just the one everything can open.

What this does to the business model

Commercial modelers historically sold three things bundled together: the modeling technology, a curated library of captures, and the plugin or pedal that plays them. TONE3000 has now made all three free at the software layer, which leaves competitors selling hardware, curation, workflow, and support. Those are real products. They are just no longer protected by the model being a secret.

Caveats before you rip out your rig

• It is 0.0.4. The repository shows 33 open issues at the time of writing. A first public release of a plugin that has to behave inside fifteen different DAWs will have edges.
• The comparison is vendor-run. See above. The methodology is sound, the incentive is obvious, and the numbers are shown as a chart.
• Quality is per-capture, not per-architecture. A library of 700,000 community uploads means a lot of them are mediocre captures of the same five amps. The architecture ceiling being high does not make a badly reamped capture good.
• Account required for the library. The plugin is MIT, the tones stream from a service with OAuth. Local .nam loading works offline, but the 700,000 number lives behind a login.

Key Takeaways

• TONE3000 shipped a free, MIT-licensed NAM plugin on September 1, 2026, in VST3, AU, AAX, CLAP, LV2 and standalone across macOS, Windows and Linux.
• It fronts a library of 700,000+ neural captures and IRs, up from 300,000 three months earlier when NAM A2 launched.
• NAM A2 is 23 layers over three dilation stacks at 48 kHz, in an 8-channel Full and a 3-channel Lite variant, and is 30-40% cheaper on CPU than the previous generation.
• A2-Lite runs at 50% CPU on a $3 ARM Cortex-M7, so the same capture file works in a DAW and inside a hardware pedal.
• The claim that A2 beats Neural DSP, ToneX and Line 6 comes from a 1,000-participant MUSHRA study that TONE3000 designed and ran itself. Sound method, interested party.
• With models, training code, inference engine and host all open, the paid amp-modeling business is now selling hardware and curation rather than a secret.

Sources: TONE3000: plugin launch announcement, TONE3000: introducing NAM Architecture 2, TONE3000: NAM A2 complete guide, GitHub: tone-3000/tone3000-plugin, Guitar.com, MusicPlayers.com, Guitar World