AEmotionStudio/moss-audio-models

MOSS-Audio 4B — MAESTRO mirror

Verbatim mirror of OpenMOSS-Team/MOSS-Audio-4B-Instruct (revision 6907a499dc0e87cc77c8ae0fe23fd0eb5476a02d, 2026-04) packaged for the MAESTRO DAW's offline runtime.

MOSS-Audio 4B-Instruct (OpenMOSS) is a general audio-understanding LLM — a Whisper-style audio encoder feeding a Qwen3-4B language model with DeepStack feature injection. It captions any recording (speech, sounds, music), transcribes speech with timestamps, reports on speakers and their emotional state, lists sound events, and answers free-form questions about the audio.

Layout

bf16/               everything MAESTRO downloads (manifest allow_patterns)
  model-0000{1..3}-of-00003.safetensors   bf16 shards, byte-identical to upstream
  model.safetensors.index.json
  config.json  generation_config.json
  tokenizer_config.json  vocab.json  merges.txt  added_tokens.json
  special_tokens_map.json  processor_config.json  chat_template.jinja
  sha256.json       per-file hashes of this folder
README.md           this file

Differences from upstream

• The remote-code Python files (configuration_moss_audio.py, processing_moss_audio.py) are not shipped: MAESTRO vendors the upstream GitHub src/ inference package (github.com/OpenMOSS/MOSS-Audio @ ce783cb7) and never uses trust_remote_code. (Upstream's weights repo ships no modeling file at all — the GitHub package is the actual torch implementation.)
• The assets/ figures and upstream README are not shipped.
• Weight shards are byte-identical to upstream (sha256-verified against the upstream LFS metadata at mirror time — see bf16/sha256.json).

License

Apache-2.0, same as upstream code and weights. Copyright the OpenMOSS team. This mirror exists so MAESTRO installs pull a pinned, verified snapshot; all credit for the model goes to its authors.