NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Mureka's music model thinks before it sings, and Kunlun says that beats Suno

Kunlun Tech's Mureka adds a planning step to AI music generation, claiming higher structure quality than rivals like Suno.

2026-11-14 · 876 words · NeuroAI
Mureka's music model thinks before it sings, and Kunlun says that beats Suno

A producer hums a hook and types a mood. Instead of rendering audio the instant the prompt lands, the model first sketches the song's skeleton — intro, verse, drop, arrangement — and then fills it in.

That pause is the whole idea behind Mureka, the music line from Shenzhen-listed Kunlun Tech (昆仑万维). Where most music models go straight from text to waveform, Mureka tries to plan before it performs.

Key takeaways

  • Kunlun Tech (昆仑万维), a Shenzhen-listed company, launched Mureka O1 and Mureka V6 in March 2025.
  • Mureka O1 is described by Kunlun as a music model using a Music Chain-of-Thought (MusiCoT) architecture, planning song structure before generating audio tokens.
  • The company said O1 surpassed Suno V4 on its evaluations and reached SOTA at release; Mureka V7 (July 2025) lifted its "good-rate" from 43.4% to 57.7%, with human-voice realism up 44%.
  • Mureka V7.5 (August 2025) improved Chinese singing via better lyric and breath detection; Mureka TTS V1 added Voice Design, and the company proposed an MoE-TTS framework.
  • Kunlun reported Mureka's annualized revenue at about US$12 million by end-March 2025, and said platform visits reached 3.33 million in March 2025, up 86.5% month-on-month.

From prompt to structure

The core claim is simple to state and hard to do: don't generate sound first, plan the music first. Kunlun says Mureka O1 uses a Music Chain-of-Thought (MusiCoT) approach that pre-generates the overall structure — sections, emotional arc, arrangement — before producing audio tokens. The bet is that songs sound more coherent when the model knows the destination before it starts the journey.

The company frames this as moving AI music "from tool to craft." Mureka V7, released in July 2025, kept the "structure first, then generate" path and added support for arbitrary-length reference audio as a style prompt, which widens how a creator can steer the result.

The benchmark it cites

Kunlun's strongest claim is that Mureka O1, at release, outperformed Suno V4 and reached the top of the field on its internal evaluations. Those are the company's own tests, and we could not independently reproduce them — but the architecture argument is separable from the scoreboard. A model that reasons about form before sound has a plausible mechanism for fewer awkward mid-song collapses.

The quality metric the company highlights is a "good-rate" (良品率): it says this rose from 43.4% on V6 to 57.7% on V7, with human-voice realism up 44% and overall audio quality nearly doubling. These figures come from Kunlun's half-year report and are self-reported rather than audited by a third party.

A music model that also speaks

Mureka is not only melody. Kunlun extended the line into voice with Mureka TTS V1, which introduced a Voice Design feature — describe a voice in text and the system generates that timbre, for real people or virtual characters. The company also proposed MoE-TTS, a Mixture-of-Experts framework for character-described speech synthesis that aims to replace fixed label-based voice control with natural-language control.

By August 2025, Mureka V7.5 targeted Chinese singing specifically, using improved speech recognition to catch lyrics and breath changes, then split phrases and pauses to make the vocal feel less mechanical. For a market where Mandarin pronunciation and tone carry meaning, that is a more decisive improvement than a generic quality bump.

The revenue question

Kunlun has been unusually specific about commercial traction. It reported Mureka's annualized revenue at about US$12 million by the end of March 2025, on roughly US$1 million in monthly run-rate, and said global visits hit 3.33 million that month, up 86.5% from February — a growth rate Kunlun describes as the fastest in the category.

Whether that pace holds is the open question. AI music sits in a crowded, litigious corner of generative media: every gain in realism sharpens the debate over training data, artist consent, and what "original" even means for a model that learned from existing songs.

Where this actually helps

Strip the hype and the useful case is narrow but real. For game studios, short-form video makers, and small brands, a model that can hum a structurally sane backing track on command removes a genuine bottleneck — the cost and delay of custom music. The "reasoning first" design matters most there, because background music fails when it drifts or repeats.

What it does not yet do is replace a writer with taste. The structure-aware model raises the floor; it does not guarantee the hook.

Honest limitations

  • Claims that Mureka surpasses Suno come from Kunlun's own evaluations, which we could not independently reproduce.
  • "Good-rate" (良品率) figures are drawn from the company's half-year report and are self-reported, not third-party audited.
  • We did not verify audio quality by listening, and "beats Suno" depends on which Suno version is the comparator.
  • This piece does not assess training-data licensing, copyright exposure, or platform availability outside China.
  • The "world's first MusiCoT music model" label is Kunlun's own claim; we could not independently verify precedence.

What readers can do now

  1. Compare Mureka's output against Suno or Udio on one short brief and judge the quality with your own ears.
  2. If you make content, use AI music for drafts, stings, and background beds rather than final masters until you verify licensing.
  3. Watch whether the "reasoning-first" design spreads to other generative media as a trend, not a one-off claim.

Related coverage

Cite this article

NeuroAI (脑机网), "Mureka's music model thinks before it sings, and Kunlun says that beats Suno," 14 November 2026, https://neuroai.site/a/na-aigc-mureka-ai-music

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories