yue vs kev
Two of the top projects in Open models, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
yue
Lyrics-to-song model that writes an editable melody and chord score first
kev
Small decision models answering yes/no, choice and score questions over text
Verdict
yue ranks higher, 80 to 78. Of the 4 score parts we measure for both, yue leads on 2, kev on 1 and 1 is even.
- Adoptionyue leads by 10 points, 76 to 66.
- FreshnessEven at 100.
- Maintenanceyue leads by 9 points, 91 to 82.
- Agent-readykev leads by 30 points, 30 to 0: a wide gap.
Pick yue if you are
Researchers and musicians running local song generation on a 24 GB NVIDIA GPU
- Editable ABC score between prompt and audio lets you inspect and change melody and chords
- Staged Python API: plan, generate_semantic, synthesize, decode; artifacts saved per run
Pick kev if you are
Engineers routing or classifying text who want a self-hosted Jev-style API
- Calibrated probabilities: each checkpoint ships with a fitted temperature
- Works with the TypeSafe Python SDK unchanged via the /v1/systemone endpoint
| What we compare | yue | kev |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 76, popular | 66, popular |
| Freshness | 100, active | 100, active |
| Maintenance | 91, healthy | 82, healthy |
| Agent-ready | 0, none | 30, minimal |
| Facts from GitHub and the README | ||
| Stars | 13.6k | 10.1k |
| License | Apache-2.0 (permissive) | Apache-2.0 (permissive) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Sep 2026 | Oct 2026 |
| Language | Python | Python |
| Docker | No | No |
| GPU | Required | Optional |
| arm64 or Apple Silicon | Mentioned | Mentioned |
yue
YuE2 generates full songs with vocals and accompaniment from lyrics and a style prompt. It first plans an editable ABC score (melody and chords), then renders semantic tokens and acoustic latents to 48 kHz stereo audio with an AR–NAR Mixture-of-Transformers and a VAE. The repo also covers zero-shot covers via SheetSage2 transcription, score editing, and an agent skill. Code is Apache-2.0; model weights are CC BY-NC 4.0 with an additional creator permission.
Who it is for: Researchers and musicians running local song generation on a 24 GB NVIDIA GPU
Strengths
- Editable ABC score between prompt and audio lets you inspect and change melody and chords
- Staged Python API: plan, generate_semantic, synthesize, decode; artifacts saved per run
- Zero-shot covers via SheetSage2 transcription with the same checkpoint, no cover fine-tuning
- Reports WildSongBench results against Suno, Mureka and others, with a reproduction guide
Weaknesses
- Needs Linux, Python 3.12 and an NVIDIA BF16 GPU with 24 GB VRAM
- Weights are CC BY-NC 4.0; commercial company use requires a separate license
- Editing renders a new recording; the original waveform is not preserved
- Benchmark numbers are automatic metrics from the authors; gaps between top means are not significant
- GPU required
- Needs Hugging Face model downloads, SheetSage2 (separate environment, for covers), PyTorch
- Models: YuE2-3B, YuE2-Vae, YuE2-Vae-legacy, SheetSage2, MERT-v2
kev
Kev is a family of four fine-tuned Qwen-based models (0.8B, 4B, 9B, 27B) that answer yes/no, multiple-choice and rating questions about a text and return calibrated probabilities. The server exposes an API matching TypeSafe's System One, so the TypeSafe Python SDK works against it. The repo includes training, evaluation and serving code, plus a Modal deploy script and fine-tuning skill.
Who it is for: Engineers routing or classifying text who want a self-hosted Jev-style API
Strengths
- Calibrated probabilities: each checkpoint ships with a fitted temperature
- Works with the TypeSafe Python SDK unchanged via the /v1/systemone endpoint
- Four sizes; 0.8B runs on a 4 GB GPU, 27B scores within a point of Jev on new sources
- Fine-tune from released checkpoints; about $1 for a Kev-4B run on a Modal H100
Weaknesses
- Not on PyPI; install from source with uv, Python 3.12 or 3.13 only
- Kev-27B is 51 GB of full weights and needs an 80 GB-class GPU
- Trails Jev on harder knowledge tasks, e.g. MMLU-Pro 0.675 vs 0.840 for Kev-27B
- Validated context is 8,192 tokens for 0.8B to 9B, despite a 65,536 server limit
- GPU optional
- Needs Modal (optional, for deploy and fine-tune skill), uv, Hugging Face Hub
- Models: Kev-0.8B, Kev-4B, Kev-9B, Kev-27B, Qwen3.5 base
- port 8009
More in Open models
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.