YuE2-3B is an open-weight ~3B-parameter song-generation model from the M-A-P (Multimodal Art Projection) research collective that plans before it renders: an AR-NAR Mixture-of-Transformers first writes a readable ABC-notation score (melody and chords), then expands it through semantic tokens into a full 48 kHz stereo song with vocals, and the same model supports zero-shot covers and score edits steered by an external LLM agent; it ships with a technical report, Hugging Face weights, and companion models (SheetSage2 transcription, MERT2). This web round fills most of the case's flagged gaps: the repo licenses code and agent skill under Apache 2.0, and third-party reporting says the weights are CC BY-NC 4.0 with a Sept 16 carve-out letting individual creators publish and sell outputs royalty-free (Noiz markets itself as the only commercially licensed official studio) — though the carve-out rests on a commit message and secondary writeup rather than read license text. Hardware claims now have partial independent support: one third-party local test reports <8GB VRAM and roughly 1× real-time generation on a 48GB card, alongside author-reported 71 seconds for a 3.6-minute song on an unquantized RTX 4090 and the audio.cpp quantized port in the case. Quality comparisons remain author-supplied: WildSongBench results vs Suno v5/v6 (best-of-8 highest reported mean) are published with full metrics and CSVs, but still describe curated self-run evaluations, not independent verification.
M-A-P independently operationalized Scott's outline-then-expand position — an editable ABC score with a stabilisation point before high-res render is Progressive Resolution and the Content Generator spike built into a frontier music model — and this round resolves the gate the case itself named: weights are CC BY-NC with an individual-creator resale carve-out (commit-message-sourced, unverified text) while consumer-GPU execution is now demonstrated (audio.cpp Q4/Q8, third-party <8GB run, Ruach on one RTX 3090). That makes YuE2 a live decision input for the ACE-Step music slot in his film pipeline — score-first songs-with-vocals on his hardware class, testable for individual use, though the noncommercial terms plus self-run Suno benchmarks mean it cannot cleanly displace MIT-licensed ACE-Step for commercial work without verifying the license text.
dev:technology.ace-stepdev:project.contentgeneratorip:framework.progressive-resolutionip:concept.two-pass-compilationdev:concept.hardware-aware-local-inferenceradar:agentic-animation-production-pattern
queries asked of Scott's wikis
- editable intermediate representation outline-then-expand Content Generator
- symbolic score ABC notation MIDI structured artifact editing
- plan-then-render two-stage generation pipeline pattern
- consumer GPU local inference VRAM budget quantization
- open-weights noncommercial license gating practical use
- LLM agent editing structured artifact agentic harness
now 0 pts/hpeak 6 pts/hcomments 0/hpeers p14momentum: steady3 platformsage 735h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
2026-10-11T03:33:19Z
A velocity spike on the Ruach Studio discussion post (existing evidence) triggered sensors, but the increase is modest engagement on an already-known implementation — no new weights, ports, license text, or quality verification. The case remains corroborated by three independent implementations (audio.cpp quantized port, firsthand MIDI-cover use, Ruach Studio RC on RTX 3090); licensing (CC BY-NC with unverified creator carve-out) and author-supplied benchmarks are still the open gates.
2026-10-07T04:20:47Z
grounded: converges/medium — M-A-P independently operationalized Scott's outline-then-expand position — an editable ABC score with a stabilisation point before high-res render is Progressiv
2026-10-07T04:09:03Z
The case crosses from plausible-release to demonstrated: independent implementations (audio.cpp dev-branch port with Q4/Q8 weights, firsthand MIDI-cover use, and the Ruach Studio rendering full score-first songs locally on one RTX 3090) substantiate downloadable weights and score-guided local execution that only the project had claimed. Licensing is now the principal unresolved gate; relevance to Scott stays an example-level echo of Content Generator.
2026-10-07T03:33:13Z
evidence attached: reddit.post.1wziir6 — First-party studio release built around YuE2 (score-first editing, LoRA training, 3090-local rendering) is concrete ecosystem adoption evidence for the case.
2026-09-20T00:23:45Z
A new firsthand account of turning existing MIDI compositions into melodically faithful covers adds practical support for score-guided control, rather than merely repeating release claims. Cross-platform attention warrants renewed interest, but the report does not establish downloadable access, reproducible local execution, or licensing.
2026-09-16T23:21:52Z
The new HN listing repeats the editable-score framing without adding release artifacts, implementation results, or licensing evidence. This remains a plausible local-model lead and an editable-intermediate-representation example, not a verified tooling opportunity for Scott.
2026-09-16T23:21:33Z
evidence attached: hn.story.49734316 — This is additional coverage of the existing YuE2 release and its editable-score approach to local music generation.
2026-09-13T06:21:40Z
The attached Reddit post repeats the open-weight release claim without supplying artifacts, licensing terms, or execution results; it is amplification, not independent corroboration. YuE2 remains an editable-score workflow example rather than a verified local-tooling opportunity for Scott.
2026-09-13T06:21:19Z
evidence attached: reddit.post.1wez03x — shared external link with case evidence
2026-09-12T00:29:30Z
The screen flags noncommercial licensing as new, but that same unsupported comment was already considered; no authoritative license or access change has arrived. The update leaves YuE2 a plausible local-model lead and editable-score workflow example, rather than a verified tooling opportunity for Scott.
2026-09-11T09:33:14Z
The refreshed discussion adds aesthetic reactions, not a new implementation, access change, or deployment result. YuE2 remains an editable-score workflow example for Scott, with reported quantized local support still awaiting inspectable performance evidence and licensing clarification.
2026-09-11T08:34:29Z
The refreshed audio.cpp comment adds a concrete RTX 5090 test setup, but the excerpt omits results and does not validate the advertised 8GB deployment path. This remains a plausible local-experimentation lead and an editable-score workflow example, not yet a demonstrated tooling opportunity for Scott.
2026-09-11T05:24:32Z
The refreshed discussion remains aesthetic reaction and repeated release claims, not a new deployment or access development. YuE2 remains an editable-intermediate-representation example for Scott; reported audio.cpp support warrants eventual verification, but practical hardware requirements and licensing remain unsettled.
2026-09-11T04:28:42Z
The refreshed comments leave the substantive picture unchanged: symbolic-score control is demonstrated, while quantized local deployment remains reported rather than independently validated. Neither aesthetic reactions nor the unsupported noncommercial-use assertion establishes a new tooling opportunity or constraint for Scott.
2026-09-11T03:24:27Z
The refreshed discussion remains amplification and aesthetic reaction, adding no independent deployment evidence or authoritative licensing clarification. Reported audio.cpp support remains a plausible local-experimentation lead, but this update does not turn the symbolic-planning example into a tooling decision for Scott.
2026-09-11T02:23:13Z
The refreshed discussion adds reactions to musical quality and a noncommercial-use assertion, not independent deployment validation or authoritative licensing terms. Reported quantized local support remains the substantive lead; the editable-score workflow still offers Scott a pattern example rather than a demonstrated tooling opportunity.
2026-09-11T01:29:32Z
Reported audio.cpp quantization support makes local experimentation plausible, but the refreshed discussion adds only an uninspected ComfyUI integration link—not independent validation of practical deployment. The symbolic-score workflow remains a useful pattern example rather than a demonstrated change to Scott's tooling decisions; licensing and ordinary-GPU performance remain unresolved.
2026-09-11T01:26:02Z
grounded: converges/low — YuE2’s editable-score-before-audio workflow independently echoes Scott’s Content Generator spike, which creates an editable outline before expanding it into con
2026-09-11T01:23:21Z
case created — The model repository and first-party demonstration establish one concrete release episode, but the supplied evidence does not establish 8GB-VRAM support or practical long-form local performance.