01.AI
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v8 · 3,479 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v8 · 3,479 words
Add missing citations, update stale details, or suggest a clearer explanation.
01.AI (Chinese: 零一万物, Língyi Wànwù) is a Chinese artificial intelligence company founded in 2023 by Kai-Fu Lee that developed the Yi family of large language models. Headquartered in Beijing, the company shipped open-source and proprietary models ranging from 6 billion to roughly one trillion parameters, and reached a valuation above $1 billion in November 2023, just eight months after Lee began building the team, making it one of the fastest startups in history to achieve unicorn status.[3][4] Its first model, the open-source Yi-34B, ranked first among all pre-trained base LLMs on the Hugging Face Open LLM Leaderboard at launch, outperforming Meta's Llama 2 70B despite having fewer than half the parameters.[1] In early 2025, 01.AI shifted its strategy away from large-scale model pre-training and toward enterprise AI solutions, partnering with Alibaba Cloud to establish a joint industrial AI laboratory.[9]
The company's Chinese name, 零一万物 (Língyi Wànwù), translates roughly to "zero-one, all things" or "zero-one, everything." The name alludes to a passage from the Tao Te Ching, the foundational Taoist philosophical text. The passage describes how all things in the universe originate from a singular source, reflecting the company's ambition to build foundational AI models that serve as the basis for a wide range of applications. The "01" in the English name also evokes binary code, the fundamental language of computing.
Kai-Fu Lee began assembling the 01.AI team in March 2023 and formally incorporated the company in May 2023. Lee, who previously served as president of Google China (2005 to 2009) and founding director of Microsoft Research China (later Microsoft Research Asia), brought decades of experience in AI research and technology leadership. Before launching 01.AI, he had been running Sinovation Ventures (创新工场), a venture capital firm he founded in 2009 that manages approximately $2.5 billion in dual-currency investment funds focused on Chinese technology companies.
Lee described large language models as a "historical opportunity" for China's technology industry.[5] Operations commenced in June 2023 with a team that grew rapidly. The founding team also included Anita Huang and Ken Qi.
By November 2023, just eight months after Lee started building the company, 01.AI had achieved unicorn status following a funding round that valued the startup at over $1 billion.[3][4] This made it one of the fastest startups in history to reach unicorn valuation.[4]
In November 2023, 01.AI released the Yi-34B and Yi-6B models, its first open-source large language models.[3] Yi-34B immediately ranked first among all open-source pre-trained base LLMs on the Hugging Face Open LLM Leaderboard, outperforming models such as Meta's Llama 2 70B and Falcon-180B on key metrics despite being significantly smaller.[1]
Throughout 2024, the company released a steady stream of models. In January 2024, multimodal vision-language models (Yi-VL-6B and Yi-VL-34B) arrived. In May 2024, the company launched both the upgraded open-source Yi-1.5 series and the proprietary Yi-Large model.[6] September 2024 brought the Yi-Coder series for programming tasks.[8] In October 2024, Yi-Lightning was released as a speed-optimized model with strong reasoning capabilities.[2]
In mid-December 2024, reports emerged that 01.AI had reassigned its pre-training algorithm and infrastructure teams. By late December, members of the pre-training team reportedly received offers from Alibaba's Tongyi (通义) division, while the infrastructure team received offers from Alibaba Cloud.[9]
In January 2025, Alibaba Cloud and 01.AI announced a joint laboratory focused on developing industrial AI models.[9] The lab combines the research and development strengths of both organizations, with Alibaba handling large-scale model training and 01.AI concentrating on smaller, cost-effective application models.[9]
Kai-Fu Lee denied rumors that 01.AI had sold its pre-training team or assets to Alibaba Cloud, calling such reports "vicious slander."[10] He explained the strategic shift by stating that "only tech giants can bear the costs of training super-large models" and that "spending on more GPUs to train large models isn't the way forward for survival."[11] He characterized large models as "teacher models" that transfer capabilities to smaller, more deployable models through techniques such as data distillation.[11]
By March 2025, 01.AI had formally stopped pre-training large language models and pivoted to selling tailored AI business solutions, using DeepSeek's open-source models as a foundation.[12] Lee publicly named DeepSeek the frontrunner in China's model race, framing China's January 2025 DeepSeek breakthrough as the country's "ChatGPT moment."[17] The company reported revenue exceeding 100 million yuan (approximately $13.7 million) in 2024, with roughly 70% from enterprise clients in gaming, finance, and energy sectors, and approximately 20 to 30% from international consumer products.[12]
Several key personnel departed during this transition. Vice President Huang Wenhao moved to ByteDance, and multimodal AI development lead Pan Xin joined another company.[12] 01.AI announced plans to spin off its gaming, finance, and other AI application units into independent businesses to pursue profitability.[12]
01.AI has raised approximately $300 million in total funding across multiple rounds.
| Round | Date | Key Investors | Valuation |
|---|---|---|---|
| Series D | November 2023 | Alibaba Cloud, Tencent, Xiaomi, Sinovation Ventures | Over $1 billion |
| Series D (continued) | August 2024 | Southeast Asian investor consortium | Not disclosed |
The November 2023 funding round attracted major Chinese technology companies as investors. Alibaba Cloud, Tencent, and Xiaomi all participated alongside Sinovation Ventures, Lee's own venture capital firm. This round propelled 01.AI to unicorn status.[3] In August 2024, a Southeast Asian consortium made an additional investment.
The Yi model family encompasses a broad set of language, vision-language, and code-generation models. The name "Yi" (一) means "one" in Chinese. The models were trained from scratch by 01.AI's research team and span open-source base models, instruction-tuned chat variants, long-context extensions, and proprietary closed-source systems.[1]
The Yi base models use a decoder-only Transformer architecture with several modern design choices:
Yi-6B uses 32 layers with a hidden size of 4,096, 32 query heads, and 4 key-value heads. Yi-34B uses 60 layers with a hidden size of 7,168, 56 query heads, and 8 key-value heads.[1]
The Yi base models were pre-trained on 3.1 trillion tokens of English and Chinese text.[1] The training corpus was carefully curated with extensive deduplication and quality filtering. For the 200K-token long-context variants, an additional 5 billion tokens of long-sequence data were used during continued pre-training. The long-context extension relied on sequence parallelism rather than sparse attention modifications.[1]
A notable aspect of Yi's development was the use of a relatively small but high-quality fine-tuning dataset. The chat models were fine-tuned on fewer than 10,000 instruction-response pairs, emphasizing data quality over quantity.[1]
The following table summarizes the major models in the Yi family:
| Model | Parameters | Release Date | Type | Context Length | License |
|---|---|---|---|---|---|
| Yi-6B | 6B | November 2023 | Base (open-source) | 4K (200K extended) | Apache 2.0 |
| Yi-34B | 34B | November 2023 | Base (open-source) | 4K (200K extended) | Apache 2.0 |
| Yi-6B-Chat | 6B | November 2023 | Chat (open-source) | 4K | Apache 2.0 |
| Yi-34B-Chat | 34B | November 2023 | Chat (open-source) | 4K | Apache 2.0 |
| Yi-VL-6B | 6B+ | January 2024 | Vision-Language (open-source) | 4K | Apache 2.0 |
| Yi-VL-34B | 34B+ | January 2024 | Vision-Language (open-source) | 4K | Apache 2.0 |
| Yi-9B | 9B | March 2024 | Base (open-source) | 4K | Apache 2.0 |
| Yi-1.5-6B | 6B | May 2024 | Base/Chat (open-source) | 4K / 16K / 32K | Apache 2.0 |
| Yi-1.5-9B | 9B | May 2024 | Base/Chat (open-source) | 4K / 16K / 32K | Apache 2.0 |
| Yi-1.5-34B | 34B | May 2024 | Base/Chat (open-source) | 4K / 16K / 32K | Apache 2.0 |
| Yi-Large | ~1T (MoE) | May 2024 | Closed-source (API) | Not disclosed | Proprietary |
| Yi-Large-Turbo | Not disclosed | May 2024 | Closed-source (API) | Not disclosed | Proprietary |
| Yi-Medium | Not disclosed | May 2024 | Closed-source (API) | 4K / 200K | Proprietary |
| Yi-Vision | Not disclosed | May 2024 | Closed-source (API, multimodal) | Not disclosed | Proprietary |
| Yi-Spark | Not disclosed | May 2024 | Closed-source (API) | Not disclosed | Proprietary |
| Yi-Coder-1.5B | 1.5B | September 2024 | Code (open-source) | 128K | Apache 2.0 |
| Yi-Coder-9B | 9B | September 2024 | Code (open-source) | 128K | Apache 2.0 |
| Yi-Lightning | Not disclosed (MoE) | October 2024 | Closed-source (API) | 64K | Proprietary |
The original Yi-34B and Yi-6B were the first models released by 01.AI.[3] Yi-34B quickly gained attention for outperforming much larger open-source models. On the Hugging Face Open LLM Leaderboard, Yi-34B ranked first among pre-trained base models as of November 2023, beating Llama 2 70B despite having fewer than half the parameters.[1]
Yi-34B achieved an MMLU score of 76.3 (5-shot), compared to 69.1 for GPT-3.5. On the Chinese C-Eval benchmark, Yi-34B scored 81.4, reflecting strong bilingual capabilities.[1] The chat variant, Yi-34B-Chat, reached second place on the AlpacaEval Leaderboard as of January 2024, trailing only GPT-4 Turbo, and achieved an Elo score of 1,110 on the LMSYS Chatbot Arena.[1]
A significant practical advantage was that Yi-34B could be quantized to 4-bit precision with less than 1% accuracy loss on MMLU/CMMLU benchmarks, enabling deployment on consumer-grade GPUs with 24 GB of memory.[1]
| Benchmark | Yi-6B | Yi-34B | GPT-3.5 | GPT-4 |
|---|---|---|---|---|
| MMLU (5-shot) | 63.2 | 76.3 | 69.1 | 83.0 |
| C-Eval (5-shot) | 72.0 | 81.4 | N/A | N/A |
| BBH (3-shot) | 42.8 | 54.3 | N/A | N/A |
| HumanEval (pass@1) | 15.9 | 23.2 | N/A | N/A |
| GSM8K | 32.5 | 67.2 | N/A | N/A |
| MATH | 4.6 | 14.4 | N/A | N/A |
Yi-9B was created through depth upscaling, a technique that duplicates middle layers (layers 12 to 28) of the Yi-6B model to expand it to 9 billion parameters. Despite this relatively simple scaling method, Yi-9B outperformed a range of similarly sized open-source models including Mistral-7B, SOLAR-10.7B, Gemma-7B, and DeepSeek-Coder-7B in code, math, commonsense reasoning, and reading comprehension tasks. It scored 39.0 on HumanEval (pass@1), roughly 2.5 times Yi-6B's score of 15.9.[1]
The Yi Vision Language models (Yi-VL-6B and Yi-VL-34B) extended the Yi chat models with visual understanding capabilities. Inspired by the open-source LLaVA architecture, these models combine a vision encoder with the Yi language model backbone. Yi-VL-34B ranked first among all open-source models on the MMMU benchmark in English and CMMMU in Chinese as of January 2024, scoring 41.6 on MMMU overall.[1]
The Yi-1.5 series, released on May 13, 2024, represented an upgraded version of the original Yi models. The upgrade involved continued pre-training on an additional 500 billion high-quality tokens and fine-tuning on 3 million diverse samples.[14] Yi-1.5 shipped in three sizes (6B, 9B, and 34B) with context length variants of 4K, 16K, and 32K tokens.[14]
Yi-1.5-34B-Chat demonstrated strong improvements over its predecessor, particularly in coding, math, reasoning, and instruction following.
| Benchmark | Yi-1.5-34B-Chat |
|---|---|
| MMLU | 76.8 |
| GSM8K | 90.2 |
| MATH | 50.1 |
| HumanEval | 75.2 |
Released on May 13, 2024, Yi-Large was 01.AI's first major closed-source model, featuring approximately one trillion parameters.[6] On Stanford's AlpacaEval 2.0 benchmark, Yi-Large achieved the highest overall Win Rate globally and the second-highest LC Win Rate (length-controlled), trailing only GPT-4 Turbo. On Chinese-language benchmarks from SuperCLUE, Yi-Large surpassed GPT-4 across six evaluation datasets.[7]
The Yi-Large API was priced at approximately 20 RMB (roughly $2.70) per million tokens, less than one-third the price of GPT-4 Turbo at the time.[6]
Alongside Yi-Large, 01.AI announced that it had begun training Yi-XLarge, a next-generation Mixture-of-Experts model at the trillion-parameter scale. Early training results showed Yi-XLarge already outperforming Yi-Large on MMLU, GPQA, HumanEval, and MATH benchmarks, putting it in competition with Claude 3 Opus and GPT-4-0409.[6] However, with the company's 2025 strategic pivot away from pre-training, it is unclear whether Yi-XLarge was ever completed.
Released on September 5, 2024, Yi-Coder is a series of open-source code language models optimized for programming tasks across 52 programming languages.[8] The series includes two parameter sizes: 1.5B and 9B, each available as both a base model and a chat variant. Yi-Coder supports a maximum context length of 128K tokens, making it suitable for repository-level code understanding.[8]
Yi-Coder-9B was trained on an additional 2.4 trillion tokens of high-quality code data sourced from GitHub repositories and code-related content filtered from Common Crawl.[15]
| Benchmark | Yi-Coder-9B-Chat |
|---|---|
| HumanEval | 85.4 |
| MBPP | 73.8 |
| LiveCodeBench | 23.4 |
Yi-Coder-9B-Chat's 23.4% pass rate on LiveCodeBench made it the only model with fewer than 10 billion parameters to exceed 20% on that benchmark, surpassing DeepSeek-Coder-33B-Instruct despite being roughly one-quarter the size.[8]
Released on October 16, 2024, Yi-Lightning was 01.AI's speed-optimized proprietary model using a Mixture-of-Experts (MoE) architecture.[2] The model introduced several technical innovations:
On its debut the model collected more than 13,000 community votes, and Yi-Lightning ranked 6th overall on the LMSYS Chatbot Arena leaderboard with an Arena score of 1,287, on par with xAI's Grok-2 and just ahead of GPT-4o-0513 (7th, 1,285).[2][18] In specialized categories, it placed 2nd in Chinese, 3rd in Multi-Turn and Math, and 4th in Coding and Hard Prompts.[2]
| Benchmark | Yi-Lightning |
|---|---|
| Chatbot Arena (Overall) | 1,287 (6th place) |
| MATH (0-shot) | 76.4% |
| HumanEval (0-shot) | 83.5% |
| GPQA (0-shot) | 50.9% |
| IFEval (Prompt Loose) | 81.9% |
| Arena-Hard | 91.8 |
| MT-Bench | 8.75 |
The training infrastructure achieved FP8 quantization performance of 1,200 TFLOPS per card on NVIDIA Hopper GPUs, with cluster management systems maintaining goodput exceeding 99%.[2]
A central element of 01.AI's approach was its commitment to open-source model releases. The code and weights for Yi-series open-source models (Yi-6B, Yi-34B, Yi-9B, Yi-1.5 variants, Yi-VL variants, and Yi-Coder variants) are distributed under the Apache 2.0 license, permitting personal, academic, and commercial use without restriction.[13]
Model weights are hosted on Hugging Face, ModelScope, and WiseModel, making them widely accessible to the global developer community.[16] The company maintained an active open-source presence on GitHub with community support through issues and a Discord server.[13]
This open-source strategy differentiated 01.AI from some Chinese competitors who used more restrictive licensing. However, the company's most capable models, including Yi-Large and Yi-Lightning, remained proprietary and accessible only through APIs on 01.AI's developer platform (platform.lingyiwanwu.com).
01.AI operates in an intensely competitive Chinese AI market. The company is often grouped with five other Chinese AI startups collectively known as the "AI Tigers" or "Six Little Tigers" (AI六小虎): Baichuan Intelligence, Zhipu AI, Moonshot AI, MiniMax, and StepFun.
These startups all emerged between 2023 and 2024 and attracted substantial investment from major Chinese technology firms. They compete with each other and with the AI divisions of large Chinese companies, most notably Alibaba (Qwen models), Baidu (Ernie), ByteDance (Doubao), and DeepSeek.
The competitive landscape shifted dramatically in January 2025 when DeepSeek released its R1 reasoning model, demonstrating frontier-level capabilities at a fraction of the training cost of Western counterparts.[12] This triggered a wave of strategic reassessments among Chinese AI startups. 01.AI was among the first to respond, halting pre-training entirely and pivoting to an application-focused business model.[12]
By contrast, Alibaba's Qwen overtook Meta's Llama as the most downloaded open-source model family on Hugging Face by September 2025, and DeepSeek continued pushing the boundaries of cost-efficient training. Chinese open-source model downloads grew from roughly 1.2% of global totals in late 2024 to approximately 30% by early 2026, a shift driven primarily by Qwen and DeepSeek.
Kai-Fu Lee (born December 3, 1961, in Taipei, Taiwan) is the founder and CEO of 01.AI. He earned a Bachelor of Science summa cum laude in computer science from Columbia University in 1983 and a Ph.D. from Carnegie Mellon University, where his doctoral research focused on speech recognition.
Lee's career in technology spans several decades and multiple major companies:
| Period | Role |
|---|---|
| 1990s | Researcher and executive at Apple and SGI |
| 1998 to 2000 | Founding Director, Microsoft Research China (later Microsoft Research Asia) |
| 2005 to 2009 | President, Google China |
| 2009 to present | Chairman and CEO, Sinovation Ventures |
| 2023 to present | Founder and CEO, 01.AI |
Lee is the author of "AI Superpowers: China, Silicon Valley, and the New World Order" (2018) and co-author of "AI 2041: Ten Visions for Our Future" (2021, with Chen Qiufan). He has served as co-chair of the World Economic Forum's Global AI Council and was named to TIME's 2023 list of the 100 Most Influential People in AI.
01.AI operates a developer platform at platform.lingyiwanwu.com (also accessible as platform.01.ai), providing API access to its proprietary models. The platform offers six model API variants optimized for different use cases:
| API Model | Primary Use Case |
|---|---|
| Yi-Large | High-capability general reasoning |
| Yi-Large-Turbo | Balanced performance and speed |
| Yi-Medium | Cost-effective general use |
| Yi-Medium-200K | Long-context processing |
| Yi-Vision | Multimodal image and text tasks |
| Yi-Spark | Lightweight, fast inference |
In March 2024, 01.AI published a comprehensive technical report titled "Yi: Open Foundation Models by 01.AI" (arXiv:2403.04652), detailing the architecture, training methodology, and benchmark evaluations of the Yi model family.[1] The paper covered the base models, chat models, long-context extensions, depth-upscaled models, and vision-language variants.[1]
In December 2024, the company released the "Yi-Lightning Technical Report" (arXiv:2412.01253), describing the MoE architecture innovations, training infrastructure optimizations, and the RAISE safety framework used in Yi-Lightning.[2]