The melody-first trick that lets Qwen-Music beat Suno on 13 metrics

Most text-to-music models try to compose the melody and the arrangement at the same time. The result often sounds crowded, the vocal line fights the instruments, the structure wanders, and nothing quite lands. Qwen-Music, the new model from Alibaba's Qwen team, takes a different route: it plans the melody first, then fills in the band around it, alongside other Alibaba audio breakthroughs like its multilingual TTS model.

Melody-first generation

The core idea is called Melody-CoT, short for chain-of-thought. Before generating the full music token sequence, the model produces an intermediate set of melody tokens that describe a coarse vocal contour, roughly 6.25 Hz, enough to preserve the shape of the melody while discarding vibrato, ornaments, and other performance details. This coarse plan acts as a structural scaffold. Once the melody is set, the model generates the full 25 Hz semantic token stream that includes instruments and arrangement. The result, according to the technical report, is better creativity, musicality, and structural coherence, similar to how Alibaba's OvisOCR2 achieves SOTA by decomposing its task.

Three-part architecture

Qwen-Music is built from three components. Qwen-Music-Tokenizer compresses audio into a single-codebook stream of music semantic tokens at 25 Hz, preserving enough information for the language model to work with. Qwen-Music-LLM, a 3-billion-parameter model initialized from Qwen3.5-Omni, generates those tokens autoregressively with the Melody-CoT mechanism built into the prediction loop. Finally, Qwen-Music-Render converts the discrete token sequence back into 48 kHz stereo waveforms through a diffusion process that enriches acoustic detail. Separating composition from rendering means the language model never has to worry about audio fidelity; it only has to get the musical structure right. Alibaba's wider AI push also includes HappyOyster, an interactive world generator that takes a different approach to generative creativity.

Quality-aware training curriculum

The training data is unusually large: more than 5 million hours of multilingual music covering hundreds of languages. The team does not just feed everything in at once. They train a separate MOS-based reward model to estimate audio quality, then rank every sample within its genre into seven quality buckets. Q1 is the best, Q7 gets discarded. Training progresses through a three-stage curriculum: first on Q3-Q6 data for broad coverage, then Q2 data for quality consolidation, and finally Q1 data for high-quality refinement. The post-training pipeline follows a similar stair-step, starting with supervised fine-tuning on curated data, then iterative DPO cycles, and finally online GSPO to refine alignment with human musical preferences. This curation approach mirrors the broader appetite for quality open models, as Ollama's recent subscription pause demonstrated.

Results and benchmarks

Tested on 600 prompts split between Chinese and English, Qwen-Music achieves state-of-the-art results on 13 of 16 objective musicality and audio-quality metrics. In blind A/B preference tests with professional human raters, it wins against Suno V5 and MiniMax Music 2.6. The model also handles cover song generation, preserving a reference melody more accurately than leading proprietary systems while allowing timbre, arrangement, and genre to change. The team extracts melody tokens from the source audio and inserts them into the prompt prefix alongside target style tags and lyrics, and the model regenerates the rest around them.

The community reaction on Hugging Face, where the paper was published, has been enthusiastic, users are asking when open weights will be released. The report does not commit to a timeline, but the open-source history of the Qwen family suggests a release is likely.

Why it matters

The Melody-CoT approach addresses a fundamental limitation of end-to-end music generation models that try to learn composition and arrangement as a single, undifferentiated problem. By decomposing the task into explicit melody planning followed by arrangement, Qwen-Music produces songs with clearer structure and more human-like musical logic. It also makes melody transfer for cover songs straightforward, since the coarse melody contour can be extracted from any audio and re-rendered in a new style.