--- title: "ACE Studio raised nearly US$40 million selling music software — then gave the model away" date: 2026-11-03 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-ace-studio-music language: en --- # ACE Studio raised nearly US$40 million selling music software — then gave the model away > ACE Studio raised nearly US$40 million selling professional music software, then released its ACE-Step models free. The paid studio is the business. A producer opens a plug-in, drags a MIDI file into a lane, and gets back a vocal take with vibrato, a breath before the chorus, and a phrasing decision nobody programmed. The model that did it is downloadable for free by anyone with a laptop. That is the odd shape of AI music right now: the tools that pay are the ones sold to people who already know what a song should sound like. ## Key takeaways - ACE Studio (ACE, from Beijing-based 时域科技) raised nearly US$40 million in a new round disclosed in September 2026 by Creation Ventures (CCV), Shunwei Capital and Alphaist. - The company says ACE Studio, its professional AI-Native music platform, now generates monthly revenue in the millions of dollars, roughly 90% of it from outside mainland China, at a gross margin of about 90%. - ACE-Step 1.5 runs on consumer hardware: under 4GB of VRAM, a full song in under two seconds on an A100 and under ten seconds on an RTX 3090, with 50-plus languages. - The same team is testing a consumer product, Miya, that turns a chat message into a sung song — an attempt to reach the mass market the professional tool does not serve. - ACE-Step 1.5 is MIT-licensed and was developed with StepFun (阶跃星辰), so the model that anchors the free tier is the one any developer can download. ## Why the professionals are the business The obvious AI-music play is mass consumption: everyone gets a song for free, the songs accumulate, and the platform that hosts them wins. ACE is betting on the opposite end of the market, and the bet has numbers attached. According to coverage in Securities Times (证券时报), China Securities Journal (中证网) and Securities Daily (证券日报), ACE Studio's monthly revenue has reached the low millions of dollars, about 90% of it from overseas markets, and the gross margin is roughly 90% because the core model is self-developed. The company is close to forty people, and the round — described as one of the largest raised by a Chinese AI music team — is earmarked for music foundation model research, global team expansion, and Miya. Two details in that disclosure matter more than the headline. First, the margin: a self-hosted model with paying professional users is a software business, not a compute giveaway. Second, the geography: with nine in ten dollars earned outside mainland China, ACE Studio is not relying on the domestic market to validate itself. That also explains the free part. ACE-Step is not a promotional sample dropped alongside a subscription. ACE-Step 1.5 is MIT-licensed, developed with StepFun, and downloadable, and the paper behind it claims that on commonly used evaluation metrics the model reaches quality beyond most commercial music models while running on a consumer GPU. The open model is the argument; the studio is the invoice. ## What a professional actually buys Ask what separates a professional music tool from a chatbot and the answer is not quality. It is control, and it is editability. ACE Studio's own product pages describe the workflow as a production surface rather than a prompt box: generate vocals and instrumentals from MIDI, then refine phonemes, arrangement, layering and video scoring. There is a plug-in for Digital Audio Workstations so the tool sits inside an existing session, and a stem splitter plus a vocal-to-MIDI path so an existing recording can be re-sung by a synthetic voice. Pre-made voices and instruments are described as royalty-free for commercial projects, while the stem separator explicitly grants no rights over third-party material. ## The open model and the honest benchmark The comparison that matters is in the ACE-Step 1.5 paper, and it is more interesting than a clean win. On objective metrics the model posts the highest AudioBox content-understanding score in the table and the second-best perceptual-quality score, and it ties for best coherence. But it does not sweep the table: Suno-v5 keeps a clear advantage on style alignment and lyric alignment. Where the authors' own blind listening test placed ACE-Step 1.5 was between two commercial versions, not above all of them. The training-data claim is the one to press on. The team says its data is entirely licensed, royalty-free or synthetic, which would be an unusually clean position in a field where most music models carry legal ambiguity. That claim is the team's, and it has not been independently verified. According to Li An, chief scientist at BrainNet (脑机网), the open release is the more consequential half of this deal: a model anyone can run locally undercuts every hosted music chatbot on cost, so the only defensible position left for a paid product is the editing surface and the rights trail. ## The consumer gambit Miya (miya.fm) is the opposite bet, and it is still a test. The product page describes a personal AI music agent you talk to: describe your night and it writes and sings the track in about a minute; hand it a 1919 recording and ask for a soul-gospel version; give it a link and it produces a diss video; hand it a group chat and it sings the conversation back. The bet is that a music agent which listens — that remembers what you replayed, what you skipped, and what mood you said you were in — changes who produces music, not just who consumes it. It is also the part most likely to collide with licensing, since remixing an existing recording on the user's behalf asks the model to do something that is not obviously covered by a licence. ## Honest limitations The financial figures here are company-supplied numbers relayed by Securities Times, China Securities Journal and Securities Daily — not audited statements, and ACE is private, so there is no filing to check them against. "Monthly revenue in the low millions" and "90% from overseas" describe a current state, not a trend line; the same coverage also reports the company is still early, with plans to keep iterating the model for the following six months. The benchmark claims in the ACE-Step 1.5 paper are the authors' own evaluation, and the training-data composition claim is a company statement rather than a published audit. The round size and investor list are consistently reported as "nearly US$40 million" from CCV, Shunwei and Alphaist — the precise amount has not been disclosed. Miya's user-facing details come from the company's own site, and no independent measurement of its output quality is available. ## What readers can do now - If you are a professional producer, price the tool against your actual session economics — hours saved per track multiplied by your own rate, against the subscription — rather than against a consumer chatbot's per-song cost. - If you plan to build on ACE-Step, read the model card and the licence terms for the training data before shipping anything commercially, and pin your version: the open model moves fast and your output is not guaranteed to be reproducible across releases. - If you are judging China's AI music sector by download charts, watch the professional-tool revenue instead. That is the number that tells you whether the vertical model business is real or whether the money is still only in trials. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-ace-studio-music Free to quote with attribution and a link to the original.