In late 2024, 01.AI (零一万物) was the poster child for China's open-weight ambition: a Beijing startup with a bilingual model family, a global leaderboard ranking, and a famous founder. Nine months later it stopped training frontier models entirely. The story is less about failure than about the economics of being a mid-sized lab in a race run by giants.
Key takeaways
- 01.AI (零一万物) was founded in March 2023 by Kai-Fu Lee (李开复), the former president of Google China, and reached unicorn status (over USD 1 billion valuation) by November 2023.
- Its Yi-34B model, released in November 2023, briefly ranked first among open-source pretrained base models on the Hugging Face Open LLM Leaderboard.
- In October 2024, Yi-Lightning placed joint third globally on the LMSYS Chatbot Arena human-preference leaderboard, behind only OpenAI and Google.
- In March 2025 the company halted pretraining of its own frontier large models (大模型) and shifted to selling enterprise AI built on third-party open weights.
- Reported 2024 revenue exceeded 100 million yuan (about USD 13.7 million, or HK$108 million), with roughly 70% from enterprise clients.
The rise: a credible open-weight challenger
01.AI's early bet was that aggressive, high-quality bilingual pretraining could compete at a fraction of the parameter count of Western open models. Yi-34B, a 34-billion-parameter model released in November 2023, supported Chinese and English and, at launch, outperformed markedly larger models on standard benchmarks. It topped the Hugging Face Open LLM Leaderboard for open pretrained base models, a real signal of community traction.
The multilingual and code-focused Yi-1.5 series followed in May 2024, and Yi-Lightning in October 2024 was the lab's strongest moment. On the LMSYS Chatbot Arena, which ranks models by human pairwise preference rather than automated metrics, Yi-Lightning landed joint third worldwide alongside xAI's Grok-2. For a Chinese startup, a top-three finish on a human-judged board was a genuine milestone.
The turn: why stop pretraining
The pivot was public and explicit. Lee stated that only the largest technology companies can absorb the cost of training super-large models, and that buying more GPUs to chase frontier scale was "not the way forward for survival." In March 2025, 01.AI formally stopped pretraining its own frontier large models (大模型).
The catalyst was structural, not a single bad quarter. As open-weight releases from better-funded labs improved, the marginal value of a startup's independently trained foundation model shrank. 01.AI chose to build on top of existing open weights, fine-tuning Chinese models such as DeepSeek, Qwen and GLM for on-premises enterprise software it calls "Boss AI." This is a defensible strategy: the differentiable value moved from the base model to integration, industry data, and deployment.
The numbers that frame the decision
01.AI reported 2024 revenue above 100 million yuan (about USD 13.7 million, or HK$108 million), with roughly 70% from enterprise clients in gaming, finance and energy. That is a real business, but far below the capital required to sustain frontier pretraining at parity with the largest labs. The comparison is not about talent; it is about the arithmetic of compute.
A 2026 signal shows the repositioning worked as a going concern. According to Bloomberg, Lee said in July 2026 that 01.AI was raising a pre-IPO round ahead of a planned Hong Kong listing in 2027, with roughly half its business already coming from outside China. That reporting, from a primary financial wire, indicates investors accepted the pivot narrative.
What the pivot tells us about model economics
The 01.AI arc is a clean illustration of a broader split in China's model industry: a small number of well-capitalized labs (and the hyperscalers) own frontier pretraining, while a large tier of capable teams moves up the stack to applications, fine-tuning, and vertical deployment. Open weights are what make that second tier possible. Without permissively licensed base models, a company like 01.AI would have had to either raise frontier-scale capital or exit. With them, it converted a research lab into a profitable application vendor.
This is not unique to 01.AI. It is the logical endpoint of cheap, capable open weights: the base model becomes infrastructure, and margin migrates to whoever controls the customer relationship and the domain data.
Honest limitations
- The revenue figure (100 million yuan / USD 13.7 million / HK$108 million) and the 70% enterprise share are reported estimates; they were not audited in this article and should be treated as indicative.
- The planned 2027 Hong Kong listing is a stated intention reported by Bloomberg, not a completed transaction; timelines for IPOs frequently shift.
- Benchmark positions (Hugging Face Open LLM Leaderboard first; LMSYS Arena joint third) were point-in-time rankings that have since been surpassed by newer models.
- 01.AI's reliance on third-party open weights means its ceiling is partly set by other labs' release cadence and license terms.
What readers can do now
- When evaluating a Chinese AI vendor, separate "we trained a foundation model" from "we deploy and fine-tune open weights" — the cost base and risk profile are completely different.
- For enterprise buyers, the 01.AI path validates a build-on-open-weights approach: negotiate for on-premises deployment and data isolation rather than custom pretraining.
- Track the Hong Kong listing filings (if they materialize) as a rare public window into the unit economics of the Chinese application-layer AI business.