『 Awesome Music Generation Models 』

A curated list of open-source music generation models for audio and song synthesis. Models are sorted by release date (newest first).

░ Table of Contents

• Music Generation Models
• Audio-Language Models (Audio Understanding)
• Additional Resources

≋≋≋≋≋≋≋≋≋≋≋≋≋≋

░ Music Generation Models

Quick Comparison

SymphonyGen

SymphonyGen

Description: A 3D hierarchical framework for contemporary cinematic orchestration, accepted at ISMIR 2026. Decomposes symphonic scores along the Bar, Track, and Event axes with a cascading decoder architecture that keeps decoding memory far below flat token streams. Conditions generation on a beat-quantized multi-voice harmony skeleton ("short-score" conditioning) that may be user-written, analyzed, or model-generated — enabling outline control while producing full orchestral textures. Refined with GRPO (Group Relative Policy Optimization) using a cross-modal acoustic reward from CLaMP 3 audio embeddings, and suppresses tonal clashes at inference time with dissonance-averse sampling.

Release Date: August 3, 2026

Features: A 3D hierarchical cascading decoder that decomposes orchestral scores along the Bar, Track, and Event axes — keeping decoding memory far below flat token streams while enabling conditioning at every structural level. A beat-quantized harmony skeleton provides "short-score" conditioning (user-written, analyzed, or model-generated), and the model is refined with GRPO against a cross-modal acoustic reward from CLaMP 3 audio embeddings, plus a dissonance-averse sampling algorithm that suppresses unintended tonal clashes during inference.

Links:
![HuggingFace][link-huggingface]
![GitHub][link-github]
![Website][link-website]
![arXiv][link-arxiv]

· · · · · · · · · · · · · ·

VocalRender

VocalRender

Description: A score-native singing voice synthesis model that directly transforms composer-oriented symbolic scores — lyrics, MIDI pitches, note values, and tempo — into expressive 48 kHz singing audio. Combines an interleaved lyric–note representation (preserving note-to-word alignment and melisma), continuous acoustic latents via AudioVAE, and autoregressive diffusion for global prosody modeling. Does not require phoneme-level durations, an explicit duration predictor, or a time-aligned acoustic reference. Fine-tuned from VoxCPM2 on the open-source CrawlSinger-OS dataset.

Release Date: July 30, 2026

Features: Score-native singing voice synthesis: an interleaved representation serializes BPM followed by each lyric syllable and its associated (pitch, note-value) pairs, retaining lyric-to-note alignment and supporting melisma — eliminating the need for phoneme-level durations or time-aligned acoustic references. AudioVAE compresses singing into continuous acoustic latents that preserve fine pitch, timbre, and articulation, while an autoregressive diffusion model handles global prosody modeling and local reconstruction.

Links:
![HuggingFace][link-huggingface]
![GitHub][link-github]
![Website][link-website]
![arXiv][link-arxiv]
![HuggingFace][link-huggingface]

· · · · · · · · · · · · · ·

Allegretto Mini

Allegretto Mini

Description: A lightweight, local-first AI music generation model by OSAMA INC (India). Synthesizes stereo audio at 44.1 kHz directly from text prompts using ONNX Runtime — no cloud API, no server-side GPU, no PyTorch dependency required. Built on a three-stage pipeline (T5Gemma text encoder → 20-layer DiT with 8-step rectified-flow sampling → SAME-S decoder) derived from the Stable Audio 3 Small architecture, independently int4-quantized and packaged for efficient local deployment. Total download size ~640 MB.

Release Date: July 27, 2026

Features: A production-grade text-to-music model compressed to ~640 MB of int4 ONNX weights that runs entirely on-device via ONNX Runtime — no Python or PyTorch required. The three-stage pipeline (T5Gemma → DiT → SAME-S) is derived from the Stable Audio 3 Small architecture but independently quantized and repackaged for CPU-first local inference, making it one of the smallest and most portable open-weight music generation models available.

Links:
![HuggingFace][link-huggingface]
![GitHub][link-github]
![Demo][link-demo]

· · · · · · · · · · · · · ·

UniMuMo

UniMuMo

Description: A self-contained safe-artifact release of UniMuMo by Motius, repackaging the authors' published weights into a single verifiable checkpoint loaded via motius.pipelines.unimumo.UniMuMoPipeline. The model is a 24-layer, 1,024-dimensional dual-stream autoregressive Transformer operating over parallel music and motion token streams at a shared 50 Hz rate (Encodec 32 kHz audio with four 2,048-entry RVQ codebooks; HumanML3D-263 motion at 60 fps; T5-base text encoder and captioner). Supported routes include text-to-motion, text-to-music, joint text-to-music-motion, music-to-dance (AIST++), dance-to-music (D2M-GAN 86-clip protocol), plus captioning on both music and motion inputs. Maximum generation length per call is 10 seconds. The upstream paper is UniMuMo: Unified Text, Music and Motion Generation (Yang et al., arXiv 2410.04534).

Release Date: July 22, 2026

Features: UniMuMo's novelty is a single dual-stream autoregressive Transformer that jointly generates synchronized music and motion tokens from text, plus the reverse routes — all from one set of weights, no external captioners or codecs required at runtime. Motius' release packages every component (Encodec, T5 encoder + captioner, SentencePiece tokenizer, HumanML3D normalization stats, configuration, provenance) into one safe artifact with explicit tensor-layout mapping and a pass-or-fail loader, so UniMuMoPipeline.from_pretrained exposes text-to-music-motion, music-to-dance, dance-to-music, and bidirectional captioning behind one entry point. Reproduction audit reports exact code/waveform parity with the upstream runtime (RMSE=0, max abs err=0) on all anchor routes.

Links:
![HuggingFace][link-huggingface]
![HuggingFace][link-huggingface]
![GitHub][link-github]
![Website][link-website]
![arXiv][link-arxiv]

· · · · · · · · · · · · · ·

Magenta RealTime 2

Magenta RealTime 2

Description: Open music generation model from Google DeepMind built for on-device streaming generation with low-latency control. Follow-up to Magenta RealTime, offering richer control and lower latency across text prompts, audio examples, and MIDI.

Release Date: May 28, 2026

Features: Decoder-only Transformer LLM that does frame-wise (not chunk-wise) autoregression over SpectroStream audio tokens, conditioned on MusicCoCa embeddings + MIDI tokens — bringing Gram-style frame-level latency to live music generation with both a base (2.4B) and small (230M) configuration sized for on-device deployment.

Links:
![HuggingFace][link-huggingface]
![GitHub][link-github]
![Blog][link-blog]
![Demo][link-demo]
![arXiv][link-arxiv]

· · · · · · · · · · · · · ·

Live Music Diffusion Models (LMDM)

Live Music Diffusion Models (LMDM)

Description: A framework for efficient fine-tuning and post-training of interactive diffusion music generators. LMDM enables text-to-music generation, live generative-delay interaction (a guitarist/saxophonist/cellist can play into the model and trigger accompaniment in real time), time-varying prompt transitions, and stem-conditioned accompaniment generation across lookahead windows (+2 s / 0 s / −2 s).

Release Date: May 21, 2026

Features: A training and post-training recipe that turns a base music-diffusion model into an interactive live instrument: efficient fine-tuning plus a generative-delay deployment with controllable accommodation across positive, zero, and negative stem lookahead — shown in production-quality demos with instrumentalists performing alongside the model in real time.

Links:
![Website][link-website]
![GitHub][link-github]
![arXiv][link-arxiv]

· · · · · · · · · · · · · ·

Stable Audio 3

Stable Audio 3

Description: A state-of-the-art open platform for fast, high-quality audio and music generation built around a diffusion transformer and the SAME (Semantic-Acoustic Music Encoder) autoencoder. Ships in four variants — Small-Music, Small-SFX, Medium, and Large (API only) — covering lightweight CPU inference up to the highest-quality 2.7B parameter model. Supports variable-length generation up to 380 seconds and built-in LoRA fine-tuning.

Release Date: May 20, 2026

Features: Open-weight diffusion transformer paired with the SAME (Semantic-Acoustic Music Encoder) autoencoder — a stereo 44.1 kHz model with 256-dimensional latents optimized for both generative tractability and high-quality reconstruction. Variable-length generation avoids wasting inference time on unused latents, and built-in stackable LoRA personalization runs at small, medium, and large scales (plus a pure-MLX LoRA path on Apple Silicon).

Links:
![GitHub][link-github]
![HuggingFace][link-huggingface]
![HuggingFace][link-huggingface]
![Demo][link-demo]
![Blog][link-blog]
![arXiv][link-arxiv]

· · · · · · · · · · · · · ·

TinyMozart v2 85M

TinyMozart v2 85M

Description: Unconditional MIDI classic piano music generator trained on the Google MAESTRO V3 MIDI dataset. Tiny (85M) and CPU-friendly, designed to make local unconditional piano generation accessible.

Release Date: May 3, 2026

Features: 85M-parameter conditional-free MIDI piano generator trained on MAESTRO V3 — designed to be the smallest viable entry point for unconditional symbolic music generation on commodity hardware, paired with a use.py inference script that runs on CPU.

Links:
![HuggingFace][link-huggingface]

· · · · · · · · · · · · · ·

ACE-Step 1.5

ACE-Step 1.5

Description: The most powerful local music generation model outperforming most commercial alternatives. Supports Mac, AMD, Intel, and CUDA devices.

Release Date: February 20, 2026