Skip to content

Model version registry

Authoritative snapshot of the model board (2026-06-08), grouped by release status. The board (Linear/ClickUp) is the live source — check there for current status; this page is the durable record of the version lineage. Data Batches and Eval Score (1–10) columns on the board were unset at snapshot.

Training (next generation — Foundation 1)

Version Family Checkpoint Description
F1.0.0 Foundation 1 Params TBD; incl. additional VST batches (incl. acoustic …)
F1.0.0A Foundation 1 Synthetic sounds; batches 5–11 just Serum sounds

F1.x is Foundation One — Deep Noise's own foundation model, the move beyond the MusicGen/Meta-checkpoint lineage below.

Testing

Version Family Checkpoint Description
0.8.1L musicgen-large Meta MusicGen large
0.8.2 musicgen-small Meta Adjusted training params — AI-assisted parameter optimisation

(The 0.8.1-Large vs 0.8.2-Small head-to-head.)

Production (current)

Version Family Checkpoint Description
0.7.1 musicgen-small Meta MusicGen Meta checkpoint with MiDasheng-enriched metadata

Retired (14 total; most recent shown — [VERIFY] the older ones)

Version Family Checkpoint Description
0.6.0 musicgen-small Meta Added Omnisphere (with REVO sounds), C3 subset
0.5.4 musicgen-small Meta Fixed piano Serum sounds without Omnisphere, C4 → C3
0.5.2 musicgen-small Meta Trained only on the instruments data corpus (showed high …)

Lineage at a glance

0.5.x → 0.6.0 → 0.7.1 (production) → 0.8.1L / 0.8.2 (testing) → F1.0.0 / F1.0.0A (training). The 0.x line is MusicGen on Meta checkpoints (musicgen-small, with a large variant at 0.8.1L); Foundation One (F1.x) is the own-model next generation. See model lineage and corpus quality.

Relationships

  • model lineage
  • training spine
  • corpus quality

Timeline

  • 2026-06-08 — Registry captured from the model board (Matt). Source: self.