--- title: "Kunlun's Mureka bets on \"reasoning\" to make AI music sound less mechanical" date: 2026-10-07 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-mureka-ai-music language: en --- # Kunlun's Mureka bets on "reasoning" to make AI music sound less mechanical > Kunlun Tech's Mureka O1 introduced Chain-of-Thought music generation and claims to outperform Suno, marking China's push into AI music as a serious product, not a gimmick. A bedroom producer has a melody in their head but no instruments, no singers and no studio. For most of AI music's short history, the songs these tools produced sounded impressively wrong — vocals that slurred, structures that wandered, mixes that flattened. On 26 March 2025, Kunlun Tech (昆仑万维), through its Skywork AI unit, launched Mureka V6 and Mureka O1, and the O1 variant made a specific claim: it is the world's first music "reasoning" model, using a Chain-of-Thought approach to plan a song's structure before generating the audio. ## What Mureka actually shipped Mureka had been live for about a year before O1 — the first model, Mureka V1 (SkyMusic), arrived in April 2024. The March 2025 release paired two things: - **Mureka V6**, the base model, which supports pure instrumental generation and AI music creation across ten languages — English, Japanese, Korean, French, Spanish, Portuguese, German, Italian, Russian and Chinese. It uses an in-context learning technique the team calls ICL to widen the soundscape and improve vocal texture and mixing. - **Mureka O1**, the headline model, which applies a music-specific Chain-of-Thought method the team named MusiCoT. Instead of generating audio token by token, it first sketches the overall musical structure, then decodes the final audio — a design meant to improve structural coherence and arrangement. Kunlun's press materials state that O1 surpassed Suno V4 on multiple metrics and that Mureka became the world's first AI music platform to open both an API and model fine-tuning. The company also said users had come from more than 100 countries and regions. ## The iteration did not stop Through 2025 the line kept improving on numbers the company itself reported. Mureka V7, released in July 2025, was positioned as moving AI music from "tool-like" to "human-like"; Kunlun said its acceptable-quality rate rose from 43.4% on V6 to 57.7% on V7, vocal realism improved 44%, and overall audio quality nearly doubled. In August 2025, Mureka V7.5 focused on Chinese songs — tone, articulation, performance technique and emotional expression. Those percentages are useful as a directional signal, not as ground truth: they are company-disclosed and benchmarked against the vendor's own criteria. What they suggest is a clear product strategy — iterate fast, localize for languages underserved by Western tools, and sell access through APIs rather than only a consumer app. ## Why "reasoning" is the interesting bet Most music models are autoregressive: they predict the next slice of audio and hope the song holds together. MusiCoT's pitch is that planning first — deciding verse, chorus, bridge and arrangement up front — produces music that hangs together as a song rather than a texture. That mirrors the broader 2025 trend of "reasoning" models in language AI being applied to creative domains. If the structure is coherent, the more surface-level flaws (a wandering bridge, a weak transition) become less fatal. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the music race is less about who generates the first catchy hook and more about who builds the licensing and rights framework that lets generated tracks be used commercially without legal friction — the model is the easy part. ## What this means for creators For independent musicians outside China, Mureka matters as a reminder that the AI music field is no longer a Suno-and-everyone-else story. A platform offering ten-language support and open fine-tuning is aimed at studios and game developers who want custom voices and backgrounds, not just novelty clips. The competitive pressure pushes all vendors toward higher vocal realism and cleaner mixes. ## Why generative media is a consumer story, not just a tech story Text-to-video, singing avatars, and AI music are arriving inside apps millions already use, so the question is no longer 'can the model do it' but 'who gets to decide what looks real.' For everyday users the practical literacy is learning to treat any polished video or voice as potentially synthetic by default. The China angle matters because some of the largest consumer deployments of these tools are happening there first. ## Honest limitations Every quantitative claim here — the Suno comparison, the 43.4% to 57.7% quality-rate jump, the 44% vocal-realism gain, the 100+ countries figure — originates from Kunlun's own press releases and financial disclosures, not from an independent audit we performed. "Quality rate" and "realism" are vendor-defined metrics, and subjective listening tests will differ by genre and taste. This article does not assess music-licensing, copyright of training data, royalty treatment, or the fine-tuning terms, all of which are decisive for commercial use and should be checked directly with the platform before any paid deployment. ## What readers can do now - Listen critically: try Mureka's free tier and compare a generated track against Suno or your current tool on the same prompt. - If you produce for a non-English market, test the language support that Western tools often neglect. - For commercial work, read the licensing and fine-tuning terms before shipping a track — the model quality is only half the story. - Treat generated vocals as a draft: human arrangement, mixing and a rights check still close the gap to release-ready music. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-mureka-ai-music Free to quote with attribution and a link to the original.