It's on the screen again, Kunlun Wanweiqi attacked the music model

Written by

Iris Vance

Updated on:July-08th-2025

It reached its peak as soon as it debuted, and this large model with pure Chinese bloodline achieved a SOTA (current best level) score.

Murek O1, a music model released by Kunlun Wanwei, has surpassed Suno, the industry leader, in the evaluation of music generation quality - equivalent to ChatGPT's dominant position in text big models - and has become another "new king" that has arrived out of the blue.

Of course, unlike ordinary large models that usually compare ability differences by the accuracy of questions, there is no completely objective evaluation standard for the quality of music content, so Kunlun Wanwei has also done its best to be fair:

In the latest music evaluation, Murek O1 demonstrated excellent music quality and surpassed Suno V4 in the final overall listening evaluation.

Figure 丨 Murekao1's overall listening experience surpasses Suno V4 in subjective evaluation

In terms of objective indicators such as pronunciation accuracy, music segment accuracy, text relevance, and production quality, Murek O1 completely outperforms Sonu V4, which currently has the highest market share. This has once again raised three soul-searching questions in the overseas AI technology community:

Figure 丨 Murek O1 scores in objective evaluation of text-generated music

Who is this? Where did he come from? How did he do it?

How to say it, the starting point of doubt is arrogance. From not seeing the rise of China's AI to not caring about breakthroughs outside of large companies, this cognitive debt will take a long time to pay off.

· · ·

As a listed company, Kunlun Wanwei has no need for the so-called "2VC" narrative. Its investment in AI research is, on the one hand, a reflection of its sensitivity to technological innovation, and on the other hand, an extension of its own business.

As early as three years ago, Kunlun Wanwei used AI-generated music and graphics to reduce costs and increase efficiency for its gaming business. The copyright of a BGM worth 20,000 yuan was replaced by an AI cost of 5 yuan. This return based on real needs objectively removed the performance nature of Kunlun Wanwei's AI map.

Zhou Yahui, the founder of Kunlun Wanwei, comments on the AI performance of various companies on WeChat every year, and often circulates golden sentences that the media likes to hear, such as "ByteDance's AI strategy in 2023 failed, but it does not affect its AI strategy in 2024 to get a full score."

When evaluating his own company, Zhou Yahui used a phrase he created: "small but beautiful" .

Kunlun Wanwei's market value is around RMB 50 billion. It obviously doesn't have much chance of winning if it really goes head-to-head with global Internet giants. But in Zhou Yahui's view, Kunlun Wanwei's progress in AI is not "small and beautiful" but "small and big and beautiful." What's the big part?

It is big in the world, and big in its position in the upstream of AI commercialization.

If you search for Mureka on YouTube, you will find that many creators are already using this product to create music. This is quite different from the diffusion path of many AI concept products - first ignited by the technology community and then looking for application scenarios - the market comes first and then "explodes".

This is related to the fact that Kunlun Wanwei's overseas business has already laid a foundation. The voice social application StarMaker is a landmark product of China's interactive entertainment going overseas. Millions of people around the world are crowded in it to sing and make charts. They are well aware of the extent to which music creators will pay for productivity.

Before this wave of AI hit, StarMaker was already building its own corpus, spending millions of dollars a month on a single small language. This accumulation is equivalent to Kunlun Wanwei's "legal plug-in" to break into the big music model today. When a high-level player appears in the novice village, any action will be a dimensionality reduction attack.

So there was a successful climb to the top for Murek O1.

· · ·

To some extent, Murek O1 will make overseas AI practitioners feel helpless and want to “stop working on it”, because it is the first large-scale music model that introduces the Chain-of-Thought.

Thinking Chain is the second evolutionary curve that OpenAI o1 and DeepSeek-R1 have brought to the large model industry. By teaching large models reasoning capabilities, it solves the problem of intelligence no longer improving after pre-training reaches a bottleneck.

However, the thought chain has almost only been used in the field of text big models and has never been attempted in the music big model. Kunlun Wanwei has achieved the goal of allowing Mureka O1 to compose music like a real singer-songwriter, using experience and thinking - rather than intuition.

In the published paper, the Mureka team realized the limitations of traditional autoregressive models when generating audio, that is, following the Transformer's prediction model, it can only spit out notes (Tokens) in sequence. After creating the thinking chain, Mureka O1 can plan and sort out the overall composition structure before generation, greatly improving the coherence of the music.

This is the deepest pain point of the current music model, no doubt about it.

Simply put, the old music model represented by Suno is prone to the characteristic of "having tune but no melody" when creating music. The presence of tune means that it can indeed be identified as a piece of music, and the lack of melody means that compared with those music actually composed by humans, the results of AI's work are not pleasant to the ear and do not have catchy artistic aesthetics.

This is consistent with the criticisms of the large text model. It seems that AI is very good at writing and can spit out words and sentences continuously, but in many cases it cannot withstand close scrutiny because there are too many traces of piling up, giving people the feeling that it has too strong an "AI flavor". More serious accusations even call AI-generated works "corpse parts."

The reasoning capability provided by Murek O1 allows AI to plan the construction process of a piece of music from scratch from a global perspective, avoiding the forced element of "taking one step at a time", which recreates a magical aesthetic foundation in actual experience.

For example, this funk-style music work "Hands up high" will definitely surprise you as much as I do. Not only is it complete, but the melody and lyrics, including the realistic vocals, almost no longer have the lingering electronic feeling when it is generated by AI, and it has reached the point where it can be released:

And the joyful country songs full of freedom:

However, for more creators who want to become famous, AI is the tool that can help them create their own identity. The rave reviews of Murek O1 are based on this strong demand, a Gutenberg-style singing equality.

After DeepSeek came out of nowhere, the domestic AI industry began to show a supply chain spillover effect similar to that in the industrial field. One example is that the large music model can learn to infer and create. More importantly, from talent density to technological breakthroughs, Chinese AI companies have begun to output, and in turn contribute their experience to the world and occupy the top positions on all the charts one by one.

Magnificent innovation, the most beautiful scenery in the history of technological development