AEmotionStudio/midashenglm-gen-models

MiDashengLM-Gen — MAESTRO mirror

Mirror of mispeech/midashenglm-gen (Xiaomi MiLM Plus, Apache-2.0, arXiv:2608.11804) maintained by AEmotionStudio for download-on-demand use inside MAESTRO.

MiDashengLM-Gen generates blended 16 kHz mono audio scenes — speech, music, sound effects and ambience in one pass — from a structured multi-view caption (<|caption|> <|asr|> <|speech|> <|sfx|> <|music|> <|env|>), 1–20 s, via Qwen3-1.7B + per-token conditional flow matching + a Vocos vocoder.

Differences from upstream

• Precision: weights may be cast fp32 → bf16 (one model.safetensors); NOTICE records the payload precision and per-file sha256 digests. Upstream fp32 was sha256-verified against the hub before casting, and a bf16 payload is published only after an A/B parity pass on GPU.
• Bundled tokenizer: tokenizer/ carries the Qwen3-1.7B tokenizer + config pinned at the revision recorded in tokenizer/tokens_added.json (upstream fetches it from the network, unpinned, at construction time). tokens_added.json also records the 28 special tokens appended in training order — loader drift fails loudly instead of generating noise.
• No remote code: the modeling_*.py trust-remote-code files are not mirrored; MAESTRO vendors the inference code (Apache-2.0) in-tree.

License

Apache-2.0 (see LICENSE). Weights © 2026 MiLM Plus Xiaomi Inc.; tokenizer files © Alibaba (Qwen), Apache-2.0. See upstream for the model card, evaluation results and use-restriction notes.