Hangry Labs YuE2 v1.0.0

Hangry Labs YuE2 v1.0.0

The first stable release of our local music studio: write lyrics, describe the
music and generate songs on your NVIDIA GPU. One Docker image includes the
browser application, native tools, HTTP APIs and MCP agent tools.

Included

• Create, Library, API and System workspaces with playback, downloads, prompt
  reuse, model status, queue information and live GPU telemetry.
• English, Norwegian Bokmål, Polish, Japanese, Simplified Chinese and Spanish
  interfaces, translated READMEs and documentation introductions.
• Native asynchronous generation jobs, OpenAI-compatible /v1/audio/music and
  /v1/audio/speech transport, and restricted MCP tools with a songwriting guide
  and one-hour download links.
• FFmpeg import/export, stereo 48 kHz audio, editable ABC scores and explicit
  token/latent replay. Eight refinement steps and two acoustic variations by
  default, with a maximum of three variations in the native studio.
• A bounded serial generation queue, cancellation, backoff information and
  disk-backed completed results with configurable retention.
• Automatic retries for malformed generated scores, pinned native dependencies,
  reusable CUDA builds and speed, VRAM and lyric-fidelity benchmark tooling.

Run it

Use the quick start for the
one-line Bash and PowerShell commands and the two model downloads. Models are
mounted read-only at /models; mount persistent output at /output.

Both registries publish the same verified top-level OCI index:

hangrylabs/yue2:v1.0.0@sha256:446bd69099513f8b1d762656d455826872b7221ab232aa05e907f3ada90c9b34
ghcr.io/hangry-labs/yue2:v1.0.0@sha256:446bd69099513f8b1d762656d455826872b7221ab232aa05e907f3ada90c9b34

After downloading both model files into data/models/, start from the directory
containing that folder.

Bash:

docker run -d --name yue2 --restart unless-stopped --gpus all -p 127.0.0.1:8087:8087 --mount "type=bind,source=$(pwd)/data/models,target=/models,readonly" --mount type=volume,source=yue2-output,target=/output hangrylabs/yue2:v1.0.0@sha256:446bd69099513f8b1d762656d455826872b7221ab232aa05e907f3ada90c9b34

PowerShell:

docker run -d --name yue2 --restart unless-stopped --gpus all -p 127.0.0.1:8087:8087 --mount "type=bind,source=$($PWD.Path)\data\models,target=/models,readonly" --mount type=volume,source=yue2-output,target=/output hangrylabs/yue2:v1.0.0@sha256:446bd69099513f8b1d762656d455826872b7221ab232aa05e907f3ada90c9b34

Open http://localhost:8087. The moving latest tag remains available for
development builds. See the attached images.json for immutable references and
validation.json for the qualification summary.

Validation and scope

Source, native CPU, UI and website checks passed for the release commit.
The fixed-seed eight-step CUDA smoke baseline generated 286.2 seconds of audio
in 86.3 seconds of pipeline time (94.1 seconds total wall time) on the tested
RTX 5070 Ti. Score, semantic tokens and acoustic latents matched the previous
qualified baseline byte-for-byte. This is one local qualification run, not a
performance guarantee. The packaged studio also generated and downloaded two
tracks while switching interface languages.

Official OpenAI-client WAV/PCM generation and the MCP skill-to-Opus-download
workflow passed. Queued and active cancellation were checked, and completed
native results and MCP links returned identical bytes after a container restart.
The packaged UI also started with networking disabled and a fresh output volume.

The stable publication passed registry parity and unauthenticated access checks
in workflow 41702398585.
Its runtime config and layers are identical to the qualified candidate; the
top-level index also includes the stable workflow's provenance and SBOM.

CUDA on Linux/amd64 is the supported distribution. Hardware qualification is
on an NVIDIA RTX 5070 Ti with 16 GB VRAM. Requests share one compute worker;
parallel generation of independent requests is not enabled. Interface language
does not translate supplied lyrics or guarantee singing-language support.
Detailed technical reference pages currently remain in English.

Models and licenses

No model weights are included in the image or release assets. Obtain the
two models separately using the documented pinned downloads. Model weights
have separate CC BY-NC 4.0 terms; application source is MIT and packaged runtime
components retain their own licenses. See the
model guide and
third-party notices.

Thanks to yue2.cpp,
MAP's YuE2, GGML and the other upstream
projects that make this studio possible.

Listen to examples ·
Documentation

GGML 0.28.0 CUDA build dependency (f196a196c06e115de93a)

Reusable GGML 0.28.0 CUDA libraries for building Yue2.

Includes the reviewed shared attention-mask broadcasting adjustment used by acoustic variations. This is a build dependency prerelease, not an application release.

• Source: official GGML 0.28.0, commit facdc44a67da17df4433adcc468b424f48e9ded2.
• Platform: Linux amd64; CUDA 12.9.1; GCC 13; C++17 dependency ABI.
• CUDA targets: 75-virtual;80-virtual;86-real;89-real;120a-real.
• Input key: f196a196c06e115de93afb2abfed0fcec0d1505c342b6d563f30c285c0efc287.
• Dependency archive SHA-256: 2c72230d0081c0be8f5dcb8d956b6356731f55deb452925412482e5025cf6411.

The dependency archive contains installed headers, static libraries, CMake exports, provenance metadata and license notices. The accompanying recipe records the matching rebuild inputs. Model weights are not included.

Earlier dependency releases remain available for rollback. Application builds select this artifact through docker/ggml-deps.json.

GGML CUDA build dependency.

ggml-deps-486cfbc9cab44f7eb35f

Reuse prebuilt GGML CUDA dependencies [skip ci]

GGML 0.28.0 CUDA build dependency (cc43a4cc92197996f17c)

Compiled official GGML 0.28.0 for the Hangry Labs YuE2 CUDA build.
This is a build dependency, not a runnable application image or a product release.

• Official source: facdc44a67da17df4433adcc468b424f48e9ded2 from ggml-org/ggml.
• Source archive SHA-256: fe81fbbfdfe89a7e11c7dd646acbbf586825e9f5cdc434ac2f9703ed3f48f10c.
• Dependency input key: cc43a4cc92197996f17c34e9a1afa09c8c1f6bb60485a6a6935a250cab23e27e.
• Platform: Linux amd64, CUDA 12.9.1 / Ubuntu 24.04, C++17, GGML_MAX_NAME=128.
• Targets: 75-virtual;80-virtual;86-real;89-real;120a-real.
• No GGML runtime source patches. No model weights, compiler or application included.
• Dependency archive: 102985070 bytes; SHA-256 30c0b1b0c506b8fa517e42d4bc1a0a3b488123512a4f7ca22b2c8e9a68aaaeee.

The library archive contains headers, static libraries, CMake exports, MIT notice
and source/toolchain/file-checksum metadata. Matching .sha256 files accompany
both archives. The small build-recipe archive contains the exact source pin,
Dockerfile, CMake configuration and verification tools, with no vendored GGML tree.

Reproduction: unpack the build-recipe archive and run
docker build --target ggml-deps --build-arg BUILD_JOBS=8 ..
The attached recipe is the authoritative build input. This dependency-only tag
anchors the existing application checkout; its automatic source-code zip does not
represent the uncommitted dependency migration. The source archive downloaded by
the recipe is pinned to the official GGML commit and checksum above.

Qualification: source-fetch CPU builds and all eight native startup checks passed;
63 tooling tests passed (2 Windows symlink skips), workflow lint and website checks
passed. The packaged full-target dependency generated 286.2 seconds of audio in
95.7 seconds of pipeline time on RTX 5070 Ti, using the fixed eight-step test1
request and context 24576. This is functional qualification on that GPU, not a
hardware support matrix, controlled performance comparison or security clearance.

The previous dependency remains unchanged for rollback:
https://github.com/hangry-labs/yue2/releases/tag/ggml-deps-486cfbc9cab44f7eb35f
Its archive SHA-256 remains
63b9621135d5de01df99ec4d20f528ca6a75f8691ef14a335ad282a65b2b9eb0.