Tencent Just Valued a 100-Person Startup at $1.42 Billion. It Has Never Pretrained a Model.

Naive AI has fewer than 100 employees, no public product, and an unresolved intellectual-property fight with its founder’s last employer. It is also, as of this month, worth $1.42 billion. The Information reported that Tencent, IDG Capital, MPCi and HSG, the firm formerly known as Sequoia Capital China, have put $400 million into the seven-month-old Beijing startup across three rounds, according to a person with direct knowledge of the financing. The company has not confirmed the numbers.

What makes the deal notable is not the size of the cheque. It is what the money is explicitly not being spent on. Naive AI does not plan to pretrain a foundation model. Instead, it is taking an existing Chinese open-weight model as a base, altering its architecture, running additional mid-training passes and applying reinforcement learning to lift performance across a range of tasks. The company has not disclosed which base model it chose. Its first release, a model simply called Naive, is expected as early as this month, and it will ship as open weights that anyone can download and modify.

The funding came in three escalating tranches: $100 million, then $180 million, then a final $120 million that closed recently. Reports in China in April put the valuation at roughly $800 million after about $300 million had been raised, so the latest round lifted the price by nearly 80 percent in five months, before a single weight has been published. Founder Dai Jifeng is an associate professor in Tsinghua University’s Department of Electronic Engineering, with earlier stints at Microsoft Research Asia and SenseTime. He argues, according to The Information, that with China’s pretraining race already dominated by DeepSeek, Moonshot, Z.ai and the big platforms, there is room for a lab whose edge is the later stages of training.

Dai’s path to Naive AI has been contested. He served as technical adviser to MiroMind, the lab founded by Shanda’s Chen Tianqiao, until January. In April he told the Washington Post that MiroMind had tried to force him to relocate overseas. MiroMind denied that account in an internal notice, said its core technology and intellectual property remained entirely its own with no third party licensed, and reserved the right to pursue civil and criminal action. No lawsuit or settlement has been publicly reported since.

The post-training economy arrives

Naive AI is the purest bet yet on a thesis that has quietly gone mainstream in 2026: the most valuable work in AI increasingly happens after pretraining. The same day The Information’s story circulated, Bloomberg reported that Harvey, the OpenAI-backed legal AI company, had moved its flagship model onto Moonshot’s Kimi K3, the 2.8-trillion-parameter open-weight model released in July, and that Abridge, Decagon and Ramp were pursuing similar pivots.

Harvey’s own research post, authored by Calvin Qi and colleagues, lays out the logic plainly. “Open-weight models have cheaper per token prices,” the team wrote. “But cost is a function of both token prices and tokens used.” By shaping rewards to favor efficient reasoning during post-training, Harvey said its Kimi K3 derivative, Tenet, completed almost twice as many held-out tasks on its Legal Agent Benchmark as the base model while holding cost flat. Training took roughly 150 Nvidia B300 GPUs over two months, a rounding error next to a frontier pretraining run. The open-weight supply chain now has enough capable bases, and enough proven downstream wins, that investors will fund specialists who never touch the bottom layer.

Why it matters

The Naive AI round reprices what counts as a model company. If a hundred researchers with a good base and a strong reinforcement-learning stack can command a unicorn valuation, then pretraining, the capital-intensive layer of the stack, looks less like the moat and more like a commodity input supplied by a handful of labs to everyone else. That is good news for the labs releasing open weights, whose models become the default substrate, and awkward news for closed-model vendors whose enterprise customers are discovering they can post-train their way off usage-based pricing.

It also exposes a fragility. Post-training inherits every limit of the base. Naive AI has not said whose model it is standing on, and until the weights ship there is no way to judge whether a $1.42 billion company can add more than a few benchmark points to someone else’s work. Valuations across China’s AI sector are already running far ahead of revenue. Rhodium Group estimated this month that Chinese AI models collectively generate about 10 percent of the annual recurring revenue of OpenAI and Anthropic, with valuation-to-ARR ratios of roughly 50 times for Moonshot and 163 times for DeepSeek. “The financing gap means it will be far more difficult for Chinese frontier AI labs to scale sustainably,” Rhodium partner Logan Wright said. “They will be heavily dependent upon a favorable climate in the equity market. Historically that’s not an easy bet in China.” Z.ai, whose shares have reportedly fallen by roughly half from their peak despite record usage, shows how fast that climate can turn.

Tencent’s presence should be read in that light. The company has been seeding nearly every credible domestic model team and runs Hunyuan in-house. A stake in a post-training specialist is a hedge on where value in the stack settles, not a verdict.

What to watch

The first Naive model is due within weeks, and its release will answer three questions at once: which Chinese open-weight base Dai picked, how much reinforcement learning and architectural surgery actually add over that base on independent benchmarks, and whether the open-weight release is complete enough for others to build on in turn. Watch too for movement in the MiroMind dispute, which could complicate a launch, and for whether the Bessent-He talks in New York produce anything that touches Chinese open-weight distribution. If Naive AI ships something that matches Harvey’s numbers, expect a wave of copycat rounds in Beijing and Shanghai. If it does not, $1.42 billion will become the sector’s cautionary tale.