Skip to content

Model lineage — from MusicGen/Meta to Foundation One

The version specifics here are confirmed against the model board (version registry, snapshot 2026-06-08). Narrative context is from the transcript evaluation.

Compiled truth

The model line has two eras:

  • MusicGen on Meta checkpoints (the 0.x line). Every released 0.x version is a MusicGen model on a Meta checkpoint — mostly musicgen-small, with a large variant at 0.8.1L. The lineage ran 0.5.x → 0.6.0 → 0.7.1 → 0.8.x, improving by changing the data each version trained on (e.g. 0.7.1 added MiDasheng-enriched metadata; 0.6.0 added Omnisphere/REVO sounds; 0.5.4 fixed piano Serum sounds; pitch handling moved C4 → C3). 0.7.1 is the current production model. 0.8.1L (large) vs 0.8.2 (small, AI-assisted parameter tuning) are the two in testing.
  • Foundation One (the F1.x line) — the own-model next generation. F1.0.0 and F1.0.0A are in training now: the move beyond MusicGen to Deep Noise's own foundation model (built on a fork of Stability AI's stable-audio-tools), trained on the ~2.5M licensed/CC0 corpus. F1.0.0A is on synthetic/Serum batches (5–11); F1.0.0 adds VST/acoustic batches.

Full table: version registry. Quality is gated by Arseniy listening to the audio samples emitted every 10 epochs — see 0006 agent system layers and quality gates.

Compute: Foundation One is migrating Google Cloud → the NVIDIA/Brev 8×H100 cluster for a 2-month sprint; indicative split 1×80 GB for the foundation model, the rest for MusicGen ([VERIFY] exact split). See training spine.

Relationships

  • version registry
  • training spine
  • training
  • corpus quality
  • README

Timeline

  • 2026-06-08 — Lineage compiled from the transcript evaluation; the speculative architecture list (ACE-step/DDSP/ASEP) was removed and the version facts reconciled to the model board — see version registry. Source: self.