Naive AI heads for an open-weight release with MiroMind's IP threat unresolved

Reports in China on April 22 put the lab at roughly $800 million after about $300 million raised ^[7]. The numbers now attributed to The Information add $100 million of capital and $620 million of valuation, about $6.20 of price for every new dollar in ^[2]. The three rounds were $100 million, then $180 million, then $120 million in the one that closed recently ^[6]. With fewer than 100 employees ^[4], that is more than $4 million raised per head ^[3].

Others in this cohort are priced higher. Lin Junyang, the former lead of Alibaba's Qwen models, closed a first round in June 2026 at a roughly $2 billion post-money valuation before releasing a product ^[9]. HSG and Gaorong put in $100 million each, and Tencent $20 million ^[9]. Tencent, IDG Capital, MPCi and HSG, the firm formerly known as Sequoia Capital China, are among the investors named in Naive AI's rounds ^[8].

The money buys a modification of someone else's model. Researchers are changing the structure of an existing Chinese open-weight system, then refining it with reinforcement learning and other processes ^[10]. The plan is to publish weights that anyone can download without charge and customize ^[11]. The report leaves the base system unnamed ^[23]. Naive AI is also researching recursive self-improvement, in which a model attempts to upgrade its own capabilities ^[12].

The ownership question turns on a license MiroMind says it never granted. Dai was technical adviser to MiroMind, the lab founded by Shanda founder Chen Tianqiao, until a joint Shanda and MiroMind statement on January 18 said he was stepping down ^[13]. MiroMind's notice says Dai first offered 15 percent of his new company for an IP license and permission to take key staff ^[14]. His investors cut that to 5 percent on March 2, the notice says, while asking for a free, perpetual, irrevocable worldwide license ^[14]. The reduced offer was a third of the original ^[4]. MiroMind also said it introduced Dai to IDG and Sequoia, reported the matter to Chinese authorities and reserved the right to pursue criminal and civil action ^[15][20].

The two accounts of the split do not match. Dai told the Washington Post on April 22 that MiroMind tried to force him to relocate overseas and that this triggered his departure ^[16]. MiroMind denied that in the April 23 notice and produced an O-1 visa approval in his name ^[17]. Dai declined to comment to Bloomberg in April ^[18].

The sourcing is thin. The price rests on one unnamed person ^[1], the allegations sit in an internal document ^[15], and five months on no filing, settlement or resolution has been reported ^[19]. That bears on the price. The provenance question stands either way. A buyer of a fine-tune needs a license chain. On this one the record has MiroMind's notice and Dai's refusal to comment.

Rhodium Group estimates published this month put the combined annual recurring revenue of Chinese AI models at about 10 percent of what OpenAI and Anthropic generate, on data available through summer 2026 ^[21]. Rhodium estimated valuation-to-ARR ratios of about 163 times for DeepSeek and 50 times for Moonshot, against 34 times for OpenAI and 21 times for Anthropic ^[22]. At OpenAI's multiple, $1.42 billion implies about $42 million of recurring revenue; at DeepSeek's, about $8.7 million ^[5]. Naive AI has not shipped its first model ^[3].

For a team deciding this quarter, the question is whether an unreleased model from a seven-month-old lab belongs on the evaluation list at all ^[5]. The longer question is who answers for what sits inside a free download. A model published as open weights comes from a party the downloader has no contract with, and if MiroMind ever files, the teams already running fine-tunes of those weights learn then what the claim covers.