A bedroom producer hums a melody into a phone, types "80s city pop, female lead, rain in the bridge," and waits. A few years ago the result was a vaguely musical noise that fell apart at the 30-second mark. The model behind the newer attempt is Mureka V8 (昆仑万维), and the difference is not louder drums — it is that the system now thinks about the song before it sings it.
Mureka, developed by Beijing-based Kunlun Tech (昆仑万维), has spent 2025 climbing from a credible music model to one that, by the company's own benchmark disclosures, sits at the top of international rankings.
Key takeaways
- Mureka V8, released January 2026, ranked first on both the Vocals and Instrumental leaderboards of Artificial Analysis, ahead of Suno V5, Google Lyria 3 Pro and MiniMax Music 2.5+.
- The underlying idea is MusiCoT (Music Chain-of-Thought): the model plans global structure — sections, emotion, arrangement — before generating audio tokens.
- A measurable quality jump came earlier: Mureka V7 (July 2025) lifted the pass-rate from 43.4% to 57.7%, with vocal realism up 44% and audio quality nearly doubled versus V6.
- V8 is described by the company as capable of producing directly commercializable finished music, not just demo clips.
The architecture bet: structure before sound
Most early music models generated audio autoregressively, predicting one token after another with limited awareness of where the piece was headed. The result was局部连贯 (locally coherent) but globally flat — intros that never resolved, bridges that wandered.
Kunlun's answer, introduced with Mureka O1 and V6 in March 2025, was MusiCoT. The name echoes the chain-of-thought technique used in language models, but applied to music: instead of diving straight into audio, the model first builds an explicit plan of the whole track. V7, shipped in July 2025, doubled down on this "先结构、后生成 (structure first, then generation)" path, pre-planning sections, emotional arcs and arrangement frameworks. The company reports that V7 surpassed Suno V4.5 on mixing quality and vocal expression, and that the pass-rate climbed from 43.4% (V6) to 57.7% while vocal realism improved 44%.
V7.5 (August 2025) targeted Chinese-language singing specifically — better breath control, phrase breaks and emotional nuance — followed by V7.6 and O2 in November 2025, which extended multilingual support. Then came V8 in January 2026, which Kunlun says comprehensively exceeds Suno V5 and can follow human-like creative logic for melody, tension and resolution. In March 2026 the company added Mureka V9, tightening text control, mixing and vocal precision.
Why a benchmark sweep matters
Artificial Analysis is a third-party evaluator, not a vendor scoreboard, which gives the V8 result more weight than a self-published demo. Kunlun states V8 took the top spot in both Vocals and Instrumental categories, beating Suno V5, Google Lyria 3 Pro and MiniMax Music 2.5+. If sustained, that positions a Chinese lab at the front of a category long associated with U.S. and European products.
The same reasoning engine spawned a speech offshoot. Mureka TTS V1 introduced Voice Design — describing a voice in text and getting that timbre back — and MoE-TTS applied a mixture-of-experts architecture to let users control voice characteristics through natural language rather than fixed tags.
What the benchmark sweep signals for China's audio labs
For most of the generative-audio boom, the headline products came from the United States and Europe. A Chinese lab reaching the top of an independent music leaderboard reframes that map. It also validates a specific technical bet: that planning a song's architecture up front, the way MusiCoT does, scales better than brute-force token prediction. Rivals have not stood still — Suno, Google and MiniMax all ship strong music models — so the ranking is a snapshot, not a coronation. But the trajectory from V6 to V8 in under a year shows how quickly the gap can close once a lab commits to a coherent architecture rather than chasing individual feature demos.
What readers can do now
- If you ship audio, run a blind A/B between Mureka V8 and your current tool. The Vocals leaderboard result is the category most creators actually hear, so test it on a real brief rather than a marketing prompt.
- Use the structure-first workflow deliberately. Feed the model a clear arrangement brief (intro-verse-chorus-bridge) instead of a one-line mood; the MusiCoT design rewards explicit structure.
- Check the API before committing. Mureka's developer platform is live, so teams can bake music generation into a product rather than pasting from a web app.
Honest limitations
- The Artificial Analysis rankings and all quality percentages are company-disclosed figures from Kunlun's annual and quarterly reports; I did not independently re-run the evaluations.
- "Surpassing Suno V5 / Lyria 3 Pro" is a comparative claim tied to that specific leaderboard snapshot; leaderboards move, and different test sets can reorder models.
- I found no independently verified public user-count or revenue figure for Mureka, so this article deliberately avoids any such number.
- Music-generation quality is subjective; a top benchmark score does not guarantee a given listener will prefer the output for a given genre.