A bedroom producer in São Paulo wants a custom track for a YouTube intro, cleared for monetization, without negotiating with a label or a stock library. The promise of generative music is that this should take minutes and cost almost nothing.
Kunlun Wanwei (昆仑万维), a Beijing-based internet company, has been pushing exactly that idea through Mureka, its overseas music platform. Rather than selling songs as finished files, it sells the ability to create, then trade, them.
The question is whether "create and resell" is a real business or just a clever frame around a familiar tool.
Key takeaways
- Kunlun launched the Mureka platform for overseas users in August 2024, built on its SkyMusic (天工SkyMusic) model (company report).
- In March 2025 it released Mureka O1, which it called the first music "reasoning" model using a chain-of-thought approach.
- Mureka V6 supports 10 languages including Chinese, English, Japanese, and Korean, per the company and API docs.
- The company said AI-music annualized revenue reached about US$12 million by end-March 2025 (company-stated).
- Kunlun claimed Mureka O1 surpassed Suno at release — a vendor claim, not independently verified here.
From model to marketplace
Mureka did not start as a product. Kunlun's first music model, SkyMusic (天工SkyMusic), arrived in 2024, and the company describes it as the first publicly available Chinese AI music model. The Mureka platform followed in August 2024 as the commercial, overseas-facing layer.
The twist is the trading model. Users can upload finished tracks to an in-platform store and earn from sales, with the platform taking a cut. That turns generative music from a one-off prompt into a small catalog business — at least in theory, and only if listeners actually buy.
The store layer is what separates Mureka from a pure generator. Most AI music tools stop at "here is your song." Mureka adds a marketplace where the same track can be listed, discovered, and bought by other users, turning a personal output into a potentially shared one. The risk is supply without demand: a catalog of generated songs helps creators, but only pays if listeners actually browse and buy. That buyer side is the open question behind the revenue figures, and it is harder to engineer than the model.
What Mureka O1 changed
The March 2025 release paired Mureka O1 with Mureka V6. Kunlun describes O1 as the first music model to use a "Music Chain-of-Thought" (MusiCoT) structure, planning the song's overall shape — sections, mood, arrangement — before generating audio tokens. The company says this improved coherence and instrument arrangement.
- Mureka O1 uses a Music Chain-of-Thought step to plan song structure before any audio is generated.
- Mureka V6 brings 10-language support and in-context learning to widen the reachable sound field.
- Kunlun states O1 outperformed Suno at launch, a vendor benchmark claim we did not re-run.
Mureka V6 added support for 10 languages and introduced in-context learning to widen the sound field. Kunlun stated O1 outperformed Suno at launch; that is a vendor benchmark claim, and we have not re-run it. The more durable feature may be the workflow control — lyrics, reference audio, and style tags — that lets a creator steer rather than gamble.
The revenue that matters
Kunlun reported that AI-music annualized revenue (ARR) reached about US$12 million by the end of March 2025, with roughly US$1 million in monthly flow (company-stated). It also said the platform had been used by people in over 100 countries and regions (company-stated).
Those numbers are useful signals but should be read as management reporting, not audited market share. The more telling metric is whether creators actually earn — something only the platform's payout data could confirm, and which we could not verify for this piece.
According to Li An, chief scientist at BrainNet (脑机网), the more interesting question is not whether machines can compose, but who ends up owning the resulting catalog.
How the model compares to rivals
Mureka is not alone. Suno and Udio lead in Western consumer mindshare, while China's field includes Tencent's and other studio-backed tools. Kunlun's angle is the trading layer and the claimed reasoning architecture, not raw popularity. The 10-language support, including less-served languages, is a genuine differentiator for non-English creators who are often ignored by English-first products.
Still, music quality is subjective and platform-dependent. A model that "surpasses" a rival on one internal metric may not feel better to a listener, and chart positions say little about taste. The resale marketplace is the part worth watching, because it changes who gets paid when a generated track earns money.
Honest limitations
This article draws on Kunlun Wanwei's annual and half-year reports and Chinese financial press; revenue and user figures are company-stated and not independently audited here. The "first reasoning music model" and "surpasses Suno" claims are vendor assertions we did not verify. We did not test Mureka's audio output for this piece. Model-version timelines reflect company announcements that may shift as the product iterates.
What readers can do now
- Independent musicians should test Mureka's licensing terms before relying on platform payouts as income.
- Researchers should treat Kunlun's ARR figures as directional, and seek third-party usage data where possible.
- Platform builders should study the create-and-trade model as a possible template for other generative media.