Baichuan Intelligence
Last reviewed
Sources
27 citations
Review status
Source-backed
Revision
v8 · 3,684 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
27 citations
Review status
Source-backed
Revision
v8 · 3,684 words
Add missing citations, update stale details, or suggest a clearer explanation.
Baichuan Intelligence (Chinese: 百川智能; pinyin: Bǎchuān Zhìnéng) is a Chinese artificial intelligence company, founded in Beijing on April 10, 2023, by former Sogou CEO Wang Xiaochuan, that builds large language models and, since 2025, specializes in medical AI. It released the open-source Baichuan and Baichuan 2 (7B and 13B) models in 2023, raised more than $1 billion in funding to reach a valuation of roughly $2.7 billion by mid-2024, and is counted among China's "Six Little Tigers" of AI.[1][2][3]
Baichuan gained rapid attention by shipping a series of open-source and closed-source models within months of its founding, with backing from Alibaba, Tencent, and Xiaomi.[2][19] After the July 2024 funding round valued it at about $2.7 billion, Baichuan briefly became China's second-most valuable AI startup, behind only Moonshot AI.[3] Since early 2025 the company has pivoted its strategic focus to healthcare AI, disbanding its finance and education teams and releasing a line of specialized medical models that culminated in the 235-billion-parameter Baichuan-M3 in January 2026.[12][13] The company's name, meaning "a hundred rivers," reflects its ambition to bring together diverse streams of knowledge into powerful AI systems.
Baichuan Intelligence was co-founded by Wang Xiaochuan and Ru Liyun on April 10, 2023, in Beijing. Wang Xiaochuan serves as the company's CEO, legal representative, and executive director, while Ru Liyun serves as company supervisor.[1] The founding came at a time of intense global interest in generative AI, following the release of ChatGPT by OpenAI in late 2022, which triggered a wave of AI startup activity across China.
Wang Xiaochuan was born in 1978 in Chengdu, Sichuan. He demonstrated exceptional aptitude in science and mathematics from a young age, winning first prize in China's National High School Mathematics Competition at age 14. In 1996, at age 17, he won a gold medal at the 8th International Olympiad in Informatics (IOI), which earned him admission to Tsinghua University.[14] Wang completed his Bachelor of Science, master's degree, and doctorate in Computer Science and Technology at Tsinghua, along with an EMBA degree.[14]
While still a student, Wang participated in the creation of the online alumni directory ChinaRen.com as a part-time technical manager. After Sohu acquired ChinaRen, Wang joined Sohu and rose through the ranks from senior technical manager to CTO. At age 27, he became the youngest Vice President at Sohu.[14]
In 2003, Wang established Sohu's R&D Center, and in 2004 he launched Sogou Search, which grew into China's second-largest search engine with over 500 million monthly active users. Wang served as Sogou's CEO from 2010 until his resignation on October 15, 2021, following Tencent Holdings' acquisition of Sogou.[14] After stepping down, Wang initially explored ventures in medicine and life sciences before founding Baichuan Intelligence in April 2023.[13]
In 2024, TIME magazine named Wang Xiaochuan to its list of the 100 Most Influential People in AI.[15]
Ru Liyun, the co-founder of Baichuan Intelligence, was previously the COO and Vice President of Sogou. He joined Sogou in 2005 and played a significant operational role in transforming the company into a major Chinese internet player. Before co-founding Baichuan, Wang and Ru had also co-founded Wuji Zhikang (Five Seasons Intelligence and Wellness) in 2022, a venture focused on healthcare and AI.[13] Ru's operational experience from scaling Sogou proved valuable in building out Baichuan's organizational infrastructure during its rapid early growth.
Baichuan Intelligence's team is composed of experienced AI professionals recruited from leading technology companies, including Sogou, Baidu, Huawei, Microsoft, ByteDance, and Tencent.[1] In July 2023, Hong Tao, former CMO of Sogou, joined as the head of commercialization efforts, though he later departed in December 2024 for personal reasons.[23] In August 2024, Professor Wen Jirong (Ji-Rong Wen), Dean of the Gaoling School of Artificial Intelligence at Renmin University of China, was appointed as Baichuan Intelligence's Chief Scientist.[16] Wen had previously spent 14 years at Microsoft Research Asia (MSRA) as a senior researcher and group manager of the Web Search and Mining Group.
As of 2024, the company employed approximately 170 people.[16]
Baichuan Intelligence has raised over $1 billion in total funding across multiple rounds, making it one of the most well-funded AI startups in China.[2] The speed at which the company attracted capital reflects both the strength of Wang Xiaochuan's reputation in the Chinese technology industry and the intense investor appetite for AI companies during 2023 and 2024.
| Round | Date | Amount | Key Investors | Post-Money Valuation |
|---|---|---|---|---|
| Angel/Seed | April 2023 | $50 million | Undisclosed | N/A |
| Series A1 | October 2023 | $300 million | Alibaba, Tencent, Xiaomi, Shunwei Capital | ~$1 billion |
| Series A | July 2024 | ~$691 million (RMB 5 billion) | Alibaba, Tencent, Xiaomi, CICC, Shenzhen Capital Group, AI industrial investment fund | ~$2.7 billion (RMB 20 billion) |
The angel round of $50 million was secured at the time of founding. Just six months later, in October 2023, Baichuan completed its A1 round of $300 million from technology giants Alibaba, Tencent, and Xiaomi, as well as Shunwei Capital (a venture capital firm chaired by Xiaomi CEO Lei Jun).[4] This round pushed Baichuan's valuation past the $1 billion mark, making it one of the fastest companies ever to achieve unicorn status.[4]
In July 2024, the company completed a larger Series A round of approximately $691 million (RMB 5 billion), reaching a valuation of roughly $2.7 billion (RMB 20 billion).[2][3] The round was supported by the same core investors along with state-backed entities such as China International Capital Corporation (CICC), the AI industrial investment fund, and Shenzhen Capital Group.[3] At that valuation Baichuan was, at the time, China's second-most valuable AI startup, trailing only Moonshot AI (valued at about $3 billion).[3] The participation of state-backed investment vehicles reflected the Chinese government's strategic interest in fostering domestic AI capabilities. Wang Xiaochuan has said the company plans to begin an IPO process in 2027.[27]
Baichuan Intelligence has released a broad range of models since its founding, progressing from smaller open-source models to large closed-source offerings and, more recently, specialized domain models for healthcare and finance. The company's model development trajectory demonstrates a clear pattern: initial open-source releases to build community trust and attract developer attention, followed by larger closed-source commercial models, and ultimately domain-specific models targeting vertical industries.
| Model | Release Date | Parameters | Training Tokens | Open Source | Key Details |
|---|---|---|---|---|---|
| Baichuan-7B | June 15, 2023 | 7 billion | 1.2 trillion | Yes | First model release; Transformer-based; 4,096 context window; scored 42.8 on C-Eval, above ChatGLM-6B's 38.9, the top native Chinese pre-trained model in its size class |
| Baichuan-13B | July 11, 2023 | 13 billion | 1.4 trillion | Yes (Apache 2.0) | Pre-training (Base) and alignment (Chat) versions; free for commercial use with license; outperformed LLaMA-13B by 40% on Chinese corpora |
| Baichuan-53B | August 8, 2023 | 53 billion | Undisclosed | No (closed-source) | First closed-source model; entered internal testing |
| Baichuan 2-7B / 13B | September 6, 2023 | 7B / 13B | 2.6 trillion | Yes | Second generation; trained on 2.6 trillion tokens (more than double Baichuan 1); uses RoPE (7B) and ALiBi (13B); includes Base and Chat variants with 4-bit quantized versions |
| Baichuan2-53B | September 25, 2023 | 53 billion | Undisclosed | No (API access) | Closed-source upgrade; API opened for enterprise customers; logical reasoning +100%, math +31% over Baichuan1-53B |
| Baichuan2-192K | October 30, 2023 | Undisclosed | Undisclosed | No (API access) | 192K token context window (approximately 350,000 Chinese characters); at the time of release, the longest context window of any LLM; 14x GPT-4's context length; achieved SOTA on 7 of 10 long-text benchmarks |
| Baichuan 3 | January 29, 2024 | 100+ billion | Undisclosed | No | Claimed to surpass GPT-4 on Chinese tasks (CMMLU, GAOKAO, AGI-Eval); breakthroughs in iterative reinforcement learning |
| Baichuan-NPC | January 9, 2024 | Undisclosed | Undisclosed | No | Specialized role-playing model for game characters; optimized for character knowledge and dialogue ability |
| Baichuan 4 | May 22, 2024 | Undisclosed | Undisclosed | No | General ability +10%, math +14%, code +9% over Baichuan 3; ranked highest on SuperCLUE Chinese benchmark; multimodal capabilities; launched alongside Baixiaoying AI assistant |
The Baichuan 2 series, documented in an academic paper published on arXiv (2309.10305) in September 2023, represents the most thoroughly documented of Baichuan's model generations.[7] Both the 7B and 13B variants were trained on 2.6 trillion tokens of multilingual data, with an expanded tokenizer vocabulary optimized for Chinese text.[7] The maximum token length was set to 32 to accommodate long Chinese phrases.
The two model sizes use different positional encoding techniques: Baichuan 2-7B employs Rotary Positional Embedding (RoPE), while Baichuan 2-13B uses ALiBi (Attention with Linear Biases), which offers improved extrapolation performance for longer sequences.[7] On public benchmarks including MMLU, CMMLU, GSM8K, and HumanEval, Baichuan 2 matched or outperformed other open-source models of similar size, with particular strength in vertical domains such as medicine and law.[7]
Starting in January 2025, Baichuan Intelligence shifted significant resources toward healthcare AI, releasing a dedicated line of medical models.[13] This medical model family represents the company's most distinctive contribution to the broader AI landscape, differentiating Baichuan from competitors who have primarily focused on general-purpose model capabilities.
| Model | Release Date | Parameters | Key Details |
|---|---|---|---|
| Baichuan-M1-preview | January 25, 2025 | Undisclosed | Deep thinking model with reasoning capabilities across language, vision, and search; first in the medical model series |
| Baichuan-M1-14B | February 2025 | 14.5 billion | Open-source medical model trained from scratch on 20 trillion tokens (14T English, 4T Chinese, 2T covering 30 languages); specialized modeling for 20+ medical departments; available on Hugging Face |
| Futang Baichuan | March 20, 2025 | Undisclosed | World's first pediatric large model; developed with Beijing Children's Hospital; covers common and complex pediatric diseases |
| Baichuan-M2 | August 2025 | 32 billion | Medical augmented reasoning model using improved GRPO algorithm; trained through multi-stage reinforcement learning; open-source (Apache 2.0); scored above 32 on HealthBench Hard, previously exceeded only by GPT-5 |
| Baichuan-M2 Plus | October 2025 | Undisclosed | Evidence-enhanced medical model with improved accuracy |
| Baichuan-M3 | January 2026 | 235 billion | Open-source multimodal medical model; conducts full clinical dialogues; ranked first on diagnosis, hallucination control, HealthBench, and HealthBench Hard, surpassing GPT-5.2; medical hallucination rate of 3.5% |
| Baichuan-M3 Plus | January 22, 2026 | Undisclosed | Upgraded version released 9 days after M3; hallucination rate reduced to 2.6%; available on Hugging Face under Apache 2.0 license |
The Baichuan-M1-14B model introduced a novel approach to medical AI by training from scratch on 20 trillion tokens rather than fine-tuning an existing general-purpose model.[11] Its architecture follows the Llama framework, incorporating pre-norm RMSNorm, SwiGLU in the feed-forward network layer, and rotary position embeddings.[25] The Baichuan-M2 model introduced a Large Verifier System that pairs a Patient Simulator, which creates realistic clinical environments from de-identified medical records, with a Clinical Rubrics Generator that dynamically produces multi-dimensional evaluation metrics.[22] On HealthBench Hard, Baichuan-M2 scored above 32, a level the paper notes had previously been exceeded only by GPT-5, while remaining open-source and far cheaper to deploy.[22] The Baichuan-M3 model, at 235 billion parameters, is notable for its ability to conduct full clinical dialogues, actively gathering patient history and making informed medical decisions in a manner similar to an experienced physician; in a pure-model setting without external tools or retrieval, its reported medical hallucination rate of 3.5% was described as the world's lowest, surpassing GPT-5.2.[12]
| Model | Release Date | Key Details |
|---|---|---|
| Baichuan4-Finance | December 2024 | Full-link financial model integrating 100+ billion pieces of bilingual financial knowledge; developed with Renmin University's School of Finance; 93.62% accuracy on FLAME-Cer benchmark, surpassing GPT-4o by nearly 20% |
| Baichuan-Audio | February 2025 | End-to-end speech interaction model; text-guided speech generation; bilingual Chinese/English real-time conversations; 8-layer RVQ audio tokenizer at 12.5 Hz frame rate; open-sourced on GitHub |
Baixiaoying is Baichuan Intelligence's consumer-facing AI assistant, launched on May 22, 2024, alongside the Baichuan 4 model.[9] Described as "an AI assistant who knows how to search," Baixiaoying differentiates itself from competitors like Baidu's Ernie Bot and Moonshot AI's Kimi through its professional search capabilities.[9] The assistant supports multimodal interaction, accepting text, image, and voice inputs.
Wang Xiaochuan has articulated a vision of developing Baixiaoying into a "super app" that touches all aspects of daily life, similar to how Tencent's WeChat combines social media, messaging, payments, and e-commerce.[9] However, the product has faced adoption challenges; reports indicate that daily active users had not exceeded 5,000 since launch.[13] In late 2025, the assistant was upgraded with integration of healthcare features as part of the company's strategic pivot.
Baichuan Intelligence provides API access to its models for enterprise customers. As of 2024, the available APIs include Baichuan 4, Baichuan3-Turbo, Baichuan3-Turbo-128k, and the Assistant API. The company's API interfaces are designed to be compatible with OpenAI's API format, allowing businesses to migrate with minimal configuration changes. The Baichuan2-53B API, launched in September 2023, marked the company's first entry into enterprise-grade commercial services.
In early 2025, Baichuan Intelligence undertook a significant strategic restructuring, disbanding its business-to-business (B2B) teams for finance and education to concentrate resources on healthcare AI.[13] This pivot was catalyzed in part by the release of DeepSeek R1 in January 2025, which disrupted the competitive landscape and forced Chinese AI companies to focus on their core strengths rather than competing across all domains.[13]
The healthcare strategy centers on the vision of "Create doctors, reform pathways, advance medicine." Wang Xiaochuan announced plans for a "Super Doctor Model" and an initiative to provide every resident of Beijing's Haidian district with a personal AI healthcare assistant.[13] The company's series of medical models, from Baichuan-M1 through Baichuan-M3, represents a systematic effort to build AI systems capable of conducting clinical-quality medical consultations.
Baichuan's partnership with Beijing Children's Hospital (affiliated with Capital Medical University) to develop the Futang Baichuan pediatric model demonstrates its approach of collaborating with leading medical institutions to build specialized, clinically validated AI tools. Wang Xiaochuan's personal interest in medicine and life sciences, which he explored after leaving Sogou, also informed the company's healthcare direction.[13]
Baichuan Intelligence is frequently listed among China's "Six Little Tigers" (六小虎) of AI, a group of prominent AI startups that emerged during the 2023 generative AI wave. The term was popularized by Chinese media, including a cover story by Caixin in January 2025.[19] The six companies are typically identified as:
| Company | Founder | Primary Focus | Notable Model |
|---|---|---|---|
| Baichuan Intelligence | Wang Xiaochuan | Healthcare AI, Chinese LLMs | Baichuan-M3 |
| Zhipu AI | Tang Jie | General-purpose AI | GLM-4 |
| MiniMax | Yan Junjie | Multimodal AI | MiniMax-01 |
| Moonshot AI (Dark Side of the Moon) | Yang Zhilin | Long-context AI | Kimi |
| 01.AI | Kai-Fu Lee | Open-source LLMs | Yi series |
| StepFun (阶跃星辰) | Jiang Daxin | Multimodal AI | Step series |
These companies collectively raised billions of dollars in 2023 and 2024, with backing from major Chinese technology firms. The first four companies in the group each reached valuations of over 20 billion yuan.[19] Zhipu AI made history in January 2026 by becoming the first large-model company to go public, listing on the Hong Kong Stock Exchange.
Baichuan faces intense competition not only from the other "Six Little Tigers" but also from the AI divisions of established Chinese technology companies, including Alibaba's Qwen (Tongyi Qianwen), Baidu's Ernie, and the independently funded DeepSeek, which has gained global attention for its cost-efficient training approaches.
Baichuan Intelligence was among the first group of AI startups to navigate China's regulatory framework for generative AI. On August 31, 2023, the company's large model services passed the filing requirements of the "Interim Measures for the Management of Generative Artificial Intelligence Services," allowing them to be opened to the public.[1] This made Baichuan one of eight companies in the first batch to receive regulatory approval for public-facing generative AI services in China.[1]
In November 2023, Baichuan Intelligence collaborated with Pengcheng Laboratory on a 128K context window model. On August 8, 2024, the company established a joint laboratory with Renmin University of China focused on large model research, with Professor Wen Jirong appointed as Baichuan's Chief Scientist.[16]
The company has published several academic papers on arXiv, including the Baichuan 2 technical report (arXiv: 2309.10305, September 2023), the Baichuan-M1 paper (arXiv: 2502.12671, February 2025), the Baichuan-Audio paper (arXiv: 2502.17239, February 2025), the Baichuan4-Finance technical report (arXiv: 2412.15270, December 2024), the Baichuan-M2 paper (arXiv: 2509.02208, September 2025), and the Baichuan-M3 paper (arXiv: 2602.06570, January 2026).[7][12][20][21][22]
Baichuan Intelligence operates through multiple entities. Beijing Baichuan Intelligence Technology Co., Ltd. was established in March 2024 with a registered capital of $50 million. Shanghai Baichuan Ronghui Technology Co., Ltd. was established in August 2024 with a registered capital of $80 million.[24]
Baichuan Intelligence has received multiple industry recognitions:
| Date | Event |
|---|---|
| April 10, 2023 | Company founded by Wang Xiaochuan and Ru Liyun with $50 million seed funding |
| June 15, 2023 | Release of Baichuan-7B, the company's first open-source model |
| July 11, 2023 | Release of Baichuan-13B with commercial use license |
| August 8, 2023 | Release of closed-source Baichuan-53B; internal testing begins |
| August 31, 2023 | Passes regulatory filing for public-facing generative AI services |
| September 6, 2023 | Release of Baichuan 2 (7B and 13B) open-source models |
| September 25, 2023 | Release of Baichuan2-53B; API opened for enterprise customers |
| October 2023 | Completes $300 million A1 funding round; achieves unicorn status |
| October 30, 2023 | Release of Baichuan2-192K with world's longest context window |
| November 2023 | Collaboration with Pengcheng Laboratory on 128K context model |
| January 9, 2024 | Release of Baichuan-NPC role-playing model |
| January 29, 2024 | Release of Baichuan 3 with 100+ billion parameters |
| May 22, 2024 | Release of Baichuan 4 and launch of Baixiaoying AI assistant |
| July 2024 | Completes ~$691 million Series A round; valuation reaches ~$2.7 billion |
| August 2024 | Joint laboratory established with Renmin University; Professor Wen Jirong appointed Chief Scientist |
| December 2024 | Release of Baichuan4-Finance model; co-founder Hong Tao departs |
| January 25, 2025 | Release of Baichuan-M1-preview medical reasoning model |
| February 2025 | Open-source release of Baichuan-M1-14B and Baichuan-Audio |
| March 2025 | Strategic restructuring to focus on healthcare; release of Futang Baichuan pediatric model |
| August 2025 | Release of open-source Baichuan-M2 (32B) medical reasoning model |
| October 2025 | Release of Baichuan-M2 Plus and upgraded Baixiaoying |
| January 2026 | Release of open-source Baichuan-M3 (235B) and Baichuan-M3 Plus medical models |