.nam file versions

As more features are added to NAM, the version of the .nam file format will be incremented. The general rules try to follow semantic versioning. That means that any changes to file contents where an older version of NeuralAmpModelerCore is either not able to understand the contents or might misunderstand them (e.g. new fields that old code is not looking for) will trigger a version bump communicating a breaking change (e.g. minor version while pre-v1.0.0; later, a major version bump). Improvements where the model will be loaded correctly, but possibly with some incomplete functionality will trigger a version bump communicating a non-breaking change (e.g. minor version or patch pre-v1.0.0).

Version history

The following table shows which versions of NeuralAmpModelerCore support which model file versions:

Core Version Support Matrix

Core Version	Latest fully-supported .nam file version
0.0.0	0.5.1
0.2.0	0.5.2
0.3.0	0.5.3
0.4.0	0.6.0
0.4.1	0.7.0

Sequential models

Sequential is an architecture-specific composition of complete child NAM models. It uses the existing top-level file envelope and does not introduce a new file version:

{
  "version": "0.7.0",
  "architecture": "Sequential",
  "config": {
    "models": [
      {"version": "0.7.0", "architecture": "WaveNet", "config": {}, "weights": [], "sample_rate": 48000},
      {"version": "0.7.0", "architecture": "Linear", "config": {}, "weights": [], "sample_rate": 48000}
    ]
  },
  "weights": [],
  "sample_rate": 48000
}

The top-level weights array is empty because the wrapper has no parameters of its own. Each entry in config.models is a complete NAM model carrying its own architecture, configuration, and weights. The top-level and child sample rates must be compatible.

Sequential files emitted by the trainer before Core support was completed used bare child configs and concatenated top-level weights. Those files omitted each child’s architecture and are not supported by this canonical format.