# Δ 2026-08-07 cycle-4 · AMT "what's NEW since cycle-3" sweep

> **Status:** This is the **persisted delta** for the 2026-08-07 cycle-4 sweep.
> Cycle-3 covered the MuScriptor hot-take + MIROS/HeartMuLa/mason369/Magenta RT2 (last commit Jul 21 2026 on muscriptor/muscriptor).
> Cycle-4 covers the **17-day post-MuScriptor window (Jul 21 → Aug 7 2026)** + a 6–12-month backfill for any 2026 releases cycle-3 didn't pick up. Notable deltas: ACE-Step v1.5 (Apache-2.0, ~2 s/song on A100), Libretto (LLM agent for symbolic music), Rubato (timestamps + score in one pass), Stable Audio 3.0 (open-weights, 6:20 music on consumer GPUs), V2N (first video-only piano transcription), and the 2025 AMT Challenge results paper.
>
> **Constraint of cycle-3 (carried forward):** the main `RESOURCES.md` and `INTEGRATION.md` are being concurrently modified by other agents; these cycle-3/cycle-4 deltas live as standalone files (`RESOURCES-cycle-3.md` / `RESOURCES-cycle-4.md`, `INTEGRATION-cycle-3.md` / `INTEGRATION-cycle-4.md`) the same way `RESOURCES-delta-cycle-2c.md` did. The hub (`index.html`) now links to both cycles.

---

## 0. What's actually new since cycle-3 (delta table)

| # | Item | Why it matters | Where it lands |
|---|---|---|---|
| 1 | **ACE-Step v1.5 (Apache-2.0)** | Apache-2.0 music-gen that runs ~2 s/song on A100, ~10 s on RTX 3090. The only Apache-2.0 song-gen that fits 8 GB VRAM. XL (4B DiT) shipped Apr 2 2026. | Watchlist for the **lyrics→song** leg of cycle-2c, alongside YuE |
| 2 | **Libretto (UC Berkeley, Jun 21 2026)** | "Giving LLM agents a sense of musical structure." LLM-native grammar for symbolic music with onset slots, voices, bar-level organization. The first agent-shaped AMT-language bridge that isn't a hack on MusicXML. | Watchlist for the **agent brain** (replaces the "no good agent-shaped AMT interface" cycle-2c open question) |
| 3 | **Rubato (arXiv:2605.24291, May 22 2026)** | Prompt-conditioned encoder-decoder that outputs **timestamped notation** in one pass (no cascaded MIDI→engraving). Beats cascades on notational accuracy. New "InterMo" text representation. | Watchlist for the **end-to-end audio→score** path. Replaces the "no notation-aware AMT" cycle-2c open question. |
| 4 | **Stable Audio 3.0 (May 20 2026)** | Open-weights Small/Medium (459M / 1.4B), 6:20 music on consumer GPUs, **trained on licensed AudioSparx + Freesound CC** — the only 2026 commercial-safe open music-gen. | Watchlist for **commercial-safe music generation** (no Sora/Suno license risk). |
| 5 | **V2N (ISMIR 2026, arXiv:2608.03419)** | First complete **video-only** piano transcription (top-view video, no audio). Onset, offset, key hold, velocity from per-frame supervision. | Long-term watch — pairs with future "self-driving lab" videos. |
| 6 | **mason369/music-to-midi — May 6 2026 release** | Direct-MIROS-weights release mirror logic, so the local MIROS path no longer breaks when the AI4Musicians mirror moves. Sister: `Mike0021/muscriptor-studio` HF Space now live. | The `setup/pipeline_starter.py` integration test in cycle-5 |
| 7 | **2025 AMT Challenge results paper (arXiv:2603.27528, Mar 29 2026)** | Full leaderboard published. MIROS = #1 (F1 0.5998), YourMT3-YPTF-MoE-M = #2, Basic Pitch = #14 (F1 0.0634, last). The 21-team, 8-valid-submission table is the **canonical 2026 AMT benchmark**. | Drop the leaderboard into `dashboard/INTEGRATION.md` §"Eval" once a year. |
| 8 | **Audiveris 5.10.2 (Mar 27 2026)** | OMR engine. Stable release. Wiki update. AGPL-3.0 — license risk for our Python venv still applies. | No change from cycle-2c |
| 9 | **Stable Audio Open Small (Mar 2026)** | On-device music gen on smartphones, <8 s for 11 s clip. **Free for commercial and non-commercial.** | Watchlist for "phone-in-the-loop" transcription preview demos. |

**Nothing removed from the existing pipeline.** The only candidate for removal is **none** — every cycle-3 entry is still the right call. The cycle-4 deltas are strictly **additions to the watchlist**, not replacements.

---

## 1. New GitHub repos / releases since cycle-3

### 1.1 Music generation — Apache-2.0 wave (the biggest gap cycle-2c flagged)

- **[ace-step/ACE-Step-1.5](https://github.com/ace-step/ACE-Step-1.5)** — `v0.1.8` (May 18 2026) · **Apache 2.0** · [arXiv:2602.00744](https://arxiv.org/abs/2602.00744)
  *A hybrid LM + Diffusion Transformer (DiT) for text/lyrics → full song. Apache 2.0 across the board. Under 2 s for a 4-minute song on A100, under 10 s on RTX 3090 — **~15× faster than LLM-only baselines**. Run on as little as 4 GB VRAM. The XL series (4B DiT) released Apr 2 2026. Three pretrained DiT variants (`base`, `sft`, `turbo`) + three LM planners (0.6B/1.7B/4B Qwen3). The first Apache-2.0 path that matches commercial-grade speed for a hobbyist rig. Replaces the cycle-2c "YuE needs 16 GB" gap for the budget-conscious user.*

- **[multimodal-art-projection/YuE](https://github.com/multimodal-art-projection/YuE)** — Apache 2.0 (relicensed Jan 30 2025) · [arXiv companion](https://github.com/multimodal-art-projection/YuE)
  *Lyrics → full song. The strongest model for anything requiring lyrics + full song structure. 6B params (model ID says "7B"), 16 GB VRAM minimum, 3-4 min output at 44.1 kHz stereo. **Apache 2.0 is real** — the relicense was a watershed moment. Re-listed here because the cycle-2c reference is now solidly Apache-2.0 across both code and weights.*

- **[stability-ai/stable-audio-3](https://huggingface.co/stability-ai/stable-audio-3)** + **[stability-ai/stable-audio-open-small](https://huggingface.co/stability-ai/stable-audio-open-small)** — open-weights Small/Medium (459M/1.4B), Small SFX, Open Small · **Stability AI Community License (free for orgs < $1 M ARR, otherwise enterprise license)**
  *Stable Audio 3.0 (May 20 2026) — variable-length music up to 6:20 on Small/Medium. Trained on **fully licensed AudioSparx + ~500k Freesound CC** recordings. The first 2026 commercial-safe open music-gen. Stable Audio Open Small (Mar 2026) is the phone-grade variant: 11 s clips in < 8 s on Arm CPU, free for commercial use. Pairs well with the cycle-3 MuScriptor audio-to-MIDI loop as the "what to transcribe" source.*

### 1.2 AMT — research + tools (cycle-3 follow-ups)

- **[Xyc-arch/Libretto](https://github.com/Xyc-arch/Libretto)** — last commit Jun 21 2026 · [arXiv:2606.22708](https://arxiv.org/abs/2606.22708)
  *Libretto — "Giving LLM Agents a Sense of Musical Structure." UC Berkeley. The first framework that gives an LLM agent a corpus-calibrated, **measurable** sense of symbolic music. LLM-native grammar with explicit onset slots, voices, bar-level organization. Agents can retrieve, generate, diagnose, and revise through musician-readable feedback. The right answer to the cycle-2c "AMT agent has no good interface" open question. Live site: https://libretto.site/.*

- **[nctamer/Rubato](https://nctamer.github.io/rubato-transcription)** (paper) + code release TBD · **research-only (check license on first commit)**
  *Rubato — "Transcribing Piano Music with Timestamps." A prompt-conditioned encoder-decoder + new "InterMo" text representation. Single autoregressive pass from audio to time-aligned score. Beats cascade (audio→MIDI→engraving) on notational accuracy. Strong on MAESTRO; trained for piano. Project page: https://nctamer.github.io/rubato-transcription. **The first "one-pass audio→sheet music" system that isn't a cascade hack.***

- **[MIROS / AI4Musicians challenge](https://www.ai4musicians.org/transcription)** — winner of the 2025 AMT Challenge (F1 0.5998 vs YourMT3-YPTF-MoE-M 0.5938) · code is in **the closed competition repos** (team-forked). The model card is at [arXiv:2603.27528](https://arxiv.org/abs/2603.27528). Public deployment is via `mason369/music-to-midi`'s `MIROS` processing mode (which ships a wrapper around the local `ai4m-miros/` checkout). The challenge winners were announced at the ICME 2025 workshop (Jun 30 2025). **The "MIROS public weights" question is still gated** — the canonical artifact is the closed team repo, not an open-weights HuggingFace model.

- **[mason369/music-to-midi](https://github.com/mason369/music-to-midi)** — last commit May 6 2026 · `v1.0.14` (May 4 2026) · MIT
  *Same project as cycle-3, but with one new piece: `download_miros_model.py` now falls back to a direct release mirror if the AI4Musicians mirror is absent (the cycle-3 install path silently broke for ~10 days). Still the closest thing to a turnkey AMT pipeline. Sister: [Mike0021/muscriptor-studio](https://huggingface.co/spaces/Mike0021/muscriptor-studio) HF Space — a fresh community MuScriptor UI using the official weights.*

- **[yonghyunk1m/V2N](https://github.com/yonghyunk1m/V2N)** — ISMIR 2026 (accepted) · [arXiv:2608.03419](https://arxiv.org/abs/2608.03419)
  *V2N — Video to Notes. **The first complete visual piano transcription (VPT) system.** From a top-view video alone, with no audio, it predicts onset, offset, key hold, and velocity for every note. Shared temporal backbone with task-specific heads, per-frame supervision. Demo: https://huggingface.co/spaces/PianoVAM/V2N. Pairs with the cycle-3 KAIST stack (D3RM + D3PIA + PianoVAM) as the "see the hands, transcribe the score" leg.*

### 1.3 Realtime / low-latency AMT (cycle-2c → cycle-4 follow-ups)

- **ISMIR 2025 real-time piano work** — paper: "Exploring System Adaptations for Minimum Latency Real-Time Piano Transcription" (arXiv:2509.07586). A baseline system that eliminates non-causal processing and shares computation across model components. The reference for "lowest possible latency" piano transcription. No new public code release in the cycle-4 window.

- **Streaming Piano Transcription with Sustain Pedal** (arXiv:2503.01362) — 380 ms latency, beat-aligned decoding. The reference for the streaming use case. No new code in the cycle-4 window.

### 1.4 Foundational / older but still relevant (carried from cycle-3, license-verified fresh)

- **[ace-step/ACE-Step](https://github.com/ace-step/ACE-Step)** — original v1.0 codebase, June 3 2025. The repo from which v1.5 branched. Same Apache-2.0 license. Kept for users who need the v1.0 weight format.

- **[lucidrains/BS-RoFormer](https://github.com/lucidrains/BS-RoFormer)** — v0.9.2 (Aug 17 2025) · MIT. The reference PyTorch implementation of the architecture powering BS-RoFormer / MelBand-RoFormer. [mlx-community](https://huggingface.co/mlx-community/mel-roformer-zfturbo-vocals-v1-mlx) added MLX ports in Apr 2026 — useful for the Apple Silicon path.

- **[AEmotionStudio/roformer-models](https://huggingface.co/AEmotionStudio/roformer-models)** — fresh stems SDR table (vocals 12.97, multistem 9.65, dereverb 19.17, denoise 27.99). Apache-2.0-compatible weights via the Mæstræa redistribution.

- **[EleutherAI/aria-amt](https://github.com/EleutherAI/aria-amt)** — Apache 2.0 · unchanged. Sister weights at [AEmotionStudio/aria-amt-models](https://huggingface.co/AEmotionStudio/aria-amt-models) (Apache 2.0 mirror of `loubb/aria-midi::piano-medium-double-1.0`).

- **[audiveris/audiveris](https://github.com/audiveris/audiveris)** — `5.10.2` (Mar 27 2026) · **AGPL-3.0**. License risk for Python venv vendoring is unchanged from cycle-2c. NoteLite (v5.11.0, Apr 29 2026) is a JDK-21 fork with Chinese UI + MIDI export.

- **[Huangkuan-Xiong/Streaming-Seq2Seq](https://github.com/Huangkuan-Xiong/Streaming-Seq2Seq)** (paper code, not always public) — for the 380 ms streaming piano paper.

### 1.5 Watchlist (gone / dormant / re-emerged)

- **MIROS public weights** — **still not on HuggingFace** as of Aug 7 2026. The closest public path is the `mason369/music-to-midi` MIROS wrapper, which ships a manually-downloaded checkpoint. Keep watching for an open release; if the AI4Musicians org posts weights, drop them into `RESOURCES.md` immediately.
- **basic-pitch** — unchanged from cycle-3 (slow maintenance but not dead; v0.4.0 is still the latest).
- **omnizart** — unchanged from cycle-3 (re-emerged under `Music-and-Culture-Technology-Lab/omnizart`; verify activity before re-listing).
- **MuScriptor PyPI** — **still "PyPI release forthcoming"** as of Aug 7 2026 (per the MuScriptor HF model card). Install path is `pip install git+https://github.com/muscriptor/muscriptor.git` or `uvx muscriptor serve`. No change.

---

## 2. New HuggingFace entries since cycle-3

### 2.1 Models (new since Jul 21 2026 / new-this-cycle)

- **[ACE-Step/ACE-Step-v1-3.5B](https://huggingface.co/ACE-Step/ACE-Step-v1-3.5B)** — 3.5 B params · **Apache 2.0** · 31 Spaces using it (last updated 2026-05)
  *The original ACE-Step DiT. The v1.5 series supersedes for new work, but the v1 weights remain canonical for pipelines that depend on the 3.5 B shape.*

- **[ACE-Step/Ace-Step1.5](https://huggingface.co/ACE-Step/Ace-Step1.5)** — v1.5 family (DiT + LM) · **MIT** (per the v1.5 README)
  *The newer v1.5 release ships under MIT (not Apache 2.0 as the v1 was). Six DiT variants (`base`, `sft`, `turbo`, `turbo-rl`, `turbo-continuous`, XL) + three LM planners (0.6B / 1.7B / 4B Qwen3). The `turbo` variant is the 8-step inference sweet spot. `turbo-rl` is the post-RL checkpoint (released later in 2026). The XL 4B DiT is the high-quality option.*

- **[stability-ai/stable-audio-3](https://huggingface.co/stability-ai/stable-audio-3)** — 1.4 B (Medium) · Stability Community License
  *Stable Audio 3.0 Medium — 6:20 music, trained on licensed data, 44.1 kHz stereo. The "Suno-safe" open alternative.*

- **[stability-ai/stable-audio-open-small](https://huggingface.co/stability-ai/stable-audio-open-small)** — small on-device model · **free for commercial + non-commercial**
  *Phone-grade text-to-audio. 11 s clips in < 8 s on Arm. Apache-style permissive.*

- **[minzwon/MusicFM](https://huggingface.co/minzwon/MusicFM)** — ICASSP 2024 · `pretrained_fma.pt` + `pretrained_msd.pt` · research-only (non-commercial)
  *Unchanged from cycle-3. Still the only piece of MIROS that is public today.*

### 2.2 Datasets / Spaces (new since Jul 21 2026)

- **[PianoVAM/V2N](https://huggingface.co/spaces/PianoVAM/V2N)** — V2N demo (ISMIR 2026). Video → notes. Live on ZeroGPU.

- **[mason369/music-to-midi](https://huggingface.co/spaces/mason369/music-to-midi)** — `d39ba9b` (May 6 2026) · Space update. The cycle-3 link is unchanged; the contents now default to MIROS / MuScriptor when those weights are present locally.

- **[Mike0021/muscriptor-studio](https://huggingface.co/spaces/Mike0021/muscriptor-studio)** — new community MuScriptor UI Space. Sister to muscriptor.kyutai.org. Good "third-party UX reference" for the watchlist.

### 2.3 Datasets (unchanged from cycle-3)

- **[projectlosangeles/maestro-v3.0.0](https://huggingface.co/datasets/projectlosangeles/maestro-v3.0.0)** — CC BY-NC-SA 4.0. v3 is **not compatible with v2** — any existing v2 checkpoint needs re-training or re-evaluation. Use the HF mirror for the lighter download; the original 101 GB ZIP is on the magenta CDN.

---

## 3. Recommended stack for the demo (cycle-4 update)

| Stage | Cycle-3 default | Cycle-4 update | Why |
|---|---|---|---|
| AMT (multi-instrument full mix) | `MuScriptor/muscriptor-medium` (HF, gated CC BY-NC 4.0) + `muscriptor/muscriptor` (GH, MIT code) | **unchanged** | Still SOTA on the authors' own test set; no cycle-4 candidate beats it on full-mix multi-instrument. |
| AMT (per-stem after separation) | `spotify/basic-pitch` (Apache-2.0) | **unchanged** | Best on a clean isolated stem. |
| AMT (piano, end-to-end with timestamps) | (n/a — cascaded) | **NEW:** Rubato (paper; code TBD) | The first non-cascaded audio→timestamped-score path. Watchlist for the "I just want the sheet music" use case. |
| AMT (video-only piano) | (n/a) | **NEW:** V2N (ISMIR 2026) | The first video-only piano transcription. Long-term watch; not in the default pipeline. |
| AMT (agent brain / symbolic interface) | `nvidia/audio-flamingo-3-hf` (Apache 2.0) | **NEW:** Libretto (UC Berkeley, arXiv:2606.22708) for the **symbolic** side | Closes the cycle-2c "no good AMT-agent interface" open question. Audio-Flamingo-3 still wins for **audio** LALM; Libretto wins for **symbolic** LALM. |
| Source separation | `StemSplitio/htdemucs-ft-onnx` (Apache-2.0) | **unchanged** | Cycle-2c's pick. No cycle-4 SOTA changes for the 4-stem case. |
| Source separation (per-vocal) | `KimberleyJensen/mel-band-roformer-vocals` | **unchanged** | Best per-vocal isolation (SDR 10.98). |
| Tempo / beat | `CPJKU/beat_this` (MIT) | **unchanged** | Only commercial-friendly SOTA beat tracker. |
| Drum | `DrumScript/DrumScript` (MIT) | **unchanged** | MIT deterministic, non-ML. |
| Lyrics | `HeartMuLa/HeartTranscriptor-oss` (Apache 2.0) | **unchanged** | The cycle-2c Apache-2.0 lyrics path. |
| **Music generation (lyrics→song, hobbyist)** | `m-a-p/YuE-s1-7B-anneal-en-cot` (Apache 2.0) | **NEW primary:** ACE-Step v1.5 (Apache 2.0) | ACE-Step fits 8 GB VRAM (YuE needs 16 GB) and runs 15× faster. Same Apache-2.0 license posture. |
| **Music generation (lyrics→song, commercial-safe)** | (n/a — Suno API only) | **NEW:** Stable Audio 3.0 Medium (Stability Community License) | Only 2026 commercial-safe open music-gen (AudioSparx + Freesound CC data). |
| **Music generation (on-device / phone)** | (n/a) | **NEW:** Stable Audio Open Small | <8 s on Arm CPU, 11 s clips, free for commercial use. |
| LALM (audio LALM) | `nvidia/audio-flamingo-3-hf` (Apache 2.0) | **unchanged** | Apache-2.0 audio LALM. |
| OMR | `liebharc/homr` (MIT) | **unchanged** | Python OMR, MIT. Beats GPL-3.0 Audiveris on Python-vendoring risk. |

**Test data:** `MusicNet` (CC-BY 4.0) for the eval, MAESTRO v3.0.0 (CC BY-NC-SA — flag for non-commercial only) for the train. Both already listed in `assets.html`.

**Why this combo (2-3 lines):** the cycle-3 baseline (MuScriptor + Basic Pitch + HeartMuLa + StemSplitio) is the right SOTA stack for the project's 4-stage AMT pipeline. Cycle-4 **adds** four long-watched items (ACE-Step, Libretto, Rubato, Stable Audio 3.0) without disrupting the existing AMT path. The biggest new shape is **ACE-Step replacing YuE** for the budget hobbyist — same license, ~⅛ the VRAM, ~15× the speed. The biggest new *capability* is **Libretto** as the first symbolic-music agent interface that actually measures what it's doing.

---

## 4. License summary (cycle-4 update)

- **Permissive (MIT / Apache 2.0 / MIT-style permissive):**
  - MuScriptor **code** (MIT)
  - heartlib + HeartMuLa (Apache 2.0 since Jan 20 2026)
  - BS-RoFormer (MIT)
  - beat_this (MIT)
  - DrumScript (MIT)
  - HOMR (MIT)
  - StemSplitio/htdemucs-ft-onnx (Apache 2.0)
  - cohere-transcribe-diarize (Apache 2.0, inherited)
  - basicpitch.cpp (MIT)
  - Magenta RealTime 2 (Apache 2.0)
  - **NEW:** ACE-Step v1 (Apache 2.0) + v1.5 (MIT)
  - **NEW:** YuE (Apache 2.0 since Jan 30 2025) — re-listed
  - **NEW:** Stable Audio Open Small (free for commercial + non-commercial)

- **Stability AI Community License** (free for orgs < $1 M ARR, paid enterprise otherwise):
  - **NEW:** Stable Audio 3.0 Small / Small SFX / Medium
  - Note: not a "real" open-source license; ship only after the org clears the ARR threshold.

- **CC BY 4.0 (commercial OK with attribution):** none new this cycle.

- **CC BY-NC 4.0 (non-commercial only — gated):**
  - MuScriptor **weights** (the model)
  - SAM-Audio weights (cycle-2c)
  - MusicFM (research-only)
  - Music Flamingo (research-only)

- **CC BY-NC-SA 4.0:**
  - MAESTRO v3.0.0 (dataset) — non-commercial only

- **GPL-3.0 / AGPL-3.0:**
  - NeuralNote (DAW plugin, GPL-3.0) — unchanged
  - Audiveris 5.10.2 (OMR, AGPL-3.0) — Python-vendoring risk unchanged

- **Research-only / unclear:**
  - MIROS (no public weights; closed challenge repo)
  - MusicFM weights (research-only)
  - Libretto (UC Berkeley; check license on first commit)
  - Rubato (paper; check license on first commit)
  - V2N (ISMIR 2026; check license on first commit)
  - **NEW:** Stable Audio 3.0 Large (API/enterprise only)

- **Watchlist (gone / dormant):** basic-pitch (slow maintenance), omnizart (re-emerged under new org — verify), MIROS public weights (still gated).

**Commercial stack assemble:** the only *fully* commercial stack today is:

```
Demucs (Apache 2.0) → Basic Pitch (Apache 2.0) per stem
+ Aria-AMT (Apache 2.0) for piano
+ HeartMuLa / HeartTranscriptor-oss (Apache 2.0) for lyrics
+ ACE-Step v1 (Apache 2.0) for music-gen
+ Stable Audio Open Small (free commercial) for on-device music-gen
+ HOMR (MIT) for OMR
```

The MuScriptor path remains the SOTA on full-mix multi-instrument, but is gated under CC BY-NC 4.0 — the commercial-only stack must use Basic Pitch per-stem as the SOTA substitute.

---

## 5. Integration with existing pipeline (see `INTEGRATION-cycle-4.md`)

The new pieces fit into the existing four-stage pipeline as follows:

| Stage | Cycle-4 change | Effect |
|---|---|---|
| Stage 1 (separation) | no change | StemSplit ONNX from cycle-2c remains default |
| Stage 2 (AMT) | no change | MuScriptor-medium → default for full-mix; Basic Pitch retained for per-stem |
| Stage 2.5 (lyrics) | no change | HeartTranscriptor-oss remains default |
| Stage 2.6 (tempo) | no change | beat_this from cycle-2c |
| Stage 3 (cleanup) | no change | music21 |
| Stage 4 (engraving) | **WATCH:** Rubato (single-pass audio→timestamped score) when code lands | The first "cascade-free" end-to-end audio→sheet. Drop into `pipeline_starter.py` as a Stage-4 alternative. |
| Eval / oracle | no change | MIROS (research-only) as the F1 oracle. 2025 AMT Challenge leaderboard is now published (arXiv:2603.27528) — the official table. |
| Agent brain (LALM) | no change | Audio Flamingo 3 from cycle-2c |
| Agent brain (symbolic) | **WATCH:** Libretto (UC Berkeley, arXiv:2606.22708) for the symbolic side | The first agent-shaped AMT interface that measures structural distance. Watch for code release. |
| Music generation (lyrics→song) | **REPLACE** YuE with **ACE-Step v1** as the budget default | Same Apache-2.0 license, ~⅛ the VRAM, ~15× the speed. YuE retained for users with 16 GB+ |
| Music generation (commercial-safe) | **ADD** Stable Audio 3.0 Medium | Only 2026 commercial-safe open music-gen. Watchlist for the "client demo that won't get sued" use case |
| Real-time / live stage | no change | NeuralNote (GPL-3.0) + RTNeural (MIT) — unchanged from cycle-2c |
| Demo (this cycle) | **REFRESH** demo/ to surface cycle-4 entries (ACE-Step, Libretto, Rubato, Stable Audio 3.0) | See `INTEGRATION-cycle-4.md` §3 |