Skip to main content
Zubnet AILearnWiki › 01.AI
Companies

01.AI

Also known as: 01.AI, Yi
A Chinese AI company founded by Kai-Fu Lee in 2023, best known for its Yi series of large language models. 01.AI combined open-weight releases such as Yi-34B, which briefly led open-model leaderboards, with larger proprietary models sold through an API. In 2025 it stepped back from training frontier base models and refocused on enterprise and consumer applications, including its Wanzhi assistant.

Why it matters

01.AI is one of the cleanest case studies in the economics of the AI boom: a startup can top open-model leaderboards, reach a reported billion-dollar valuation in its first year, and still conclude that training frontier models is a bad business. Its Yi models remain in circulation, and its 2025 pivot foreshadowed the consolidation pressure now felt by every mid-sized model lab.

Deep Dive

Kai-Fu Lee — the former head of Google China and founder of the Sinovation Ventures fund — started 01.AI in 2023 with a two-track plan: train Chinese frontier models and build products on top of them. The company, whose Chinese name Lingyi Wanwu plays on the words for “zero-one,” moved unusually fast, releasing its first open models within months of founding. For a while it looked like China's answer to the Western foundation-model labs: the Yi series sat near the top of the Hugging Face open-model leaderboards, and the company ran an open-weight line and a proprietary API business side by side. Roughly two years in, it became one of the first prominent labs to publicly walk away from the pretraining race.

The Yi Model Family

The Yi series began in November 2023 with Yi-6B and Yi-34B, released as open weights under a permissive license. Yi-34B debuted at the top of the Hugging Face Open LLM Leaderboard for pretrained models — a first for a Chinese lab in terms of visibility with Western developers — and a 200K-token context window variant followed for long-document work. The family broadened through 2024 with the Yi-1.5 refresh, Yi-VL multimodal vision-language models, and Yi-Coder for programming, while the proprietary side offered larger closed models such as Yi-Large and Yi-Lightning through an API. Running an open track for adoption and a closed track for revenue at the same time was the company's signature structure. Early scrutiny also found that Yi-34B initially reused Meta's Llama architecture without attribution in its configuration files; 01.AI corrected the attribution, and the episode became a standard citation in arguments about what open releases owe their predecessors.

Open Weights as a Market Strategy

Releasing the Yi models openly was a deliberate go-to-market move rather than pure altruism. A new lab with no developer mindshare gets attention, benchmarks, and integrations for free when anyone can download its models, and every fine-tune and deployment becomes quiet marketing for the paid API. This put 01.AI in direct competition with Meta's Llama and Alibaba's Qwen for the default-choice slot among developers — a slot Qwen largely came to occupy. The strategy has real limits: downloads do not convert into revenue on their own, and every improvement to the open model can undercut the proprietary one. That tension between open adoption and paid differentiation runs through 01.AI's whole story.

The Pivot Wasn't a Failure

The common reading of 01.AI's 2025 turn is that it lost the model race and quit. What actually happened is more specific: pretraining competitive frontier base models costs on the order of hundreds of millions of dollars per generation, and DeepSeek's rise showed that strong open models were becoming a commodity whose value few companies would capture. 01.AI stopped pretraining its largest models, deepened its relationship with Alibaba Cloud — including a reported joint-lab arrangement — and redirected effort toward applications such as Wanzhi, an AI productivity assistant for enterprise and consumer users, built on whichever models make economic sense. Treating that as failure misses the point: it is the rational response to brutal pretraining economics, and other mid-sized labs have since made similar calls. The misconception persists because the industry spent years measuring labs by leaderboard position rather than by whether the business works.

The China Cohort

01.AI is usually discussed alongside the group of startups often called China's “AI tigers”: Zhipu AI, Moonshot AI, MiniMax, and StepFun, plus larger players such as Alibaba, Baidu, and ByteDance, and later DeepSeek. Each found a different answer to the same question of how to survive next to giants: Zhipu leaned into government and enterprise contracts, Moonshot into consumer chat, and DeepSeek into research-first open models. 01.AI's answer — open models for credibility, then applications for revenue — is one of the cleaner articulations of the trade. The cohort matters beyond China because its price wars and open-weight norms have repeatedly spilled into the global market, and the Yi models were an early example of a Chinese release resetting expectations for what a free model could do.

← All Terms
ESC