Baichuan

private

china · Founded 2023-04 · Valuation: $2.7B as of 2024-07

Baichuan Intelligence (百川智能) was founded in April 2023 by Wang Xiaochuan, the former CEO of Sogou (acquired by Tencent for $3.5B in 2021). Wang is one of the most prominent tech entrepreneurs in China, having built Sogou from a search startup into a major internet platform.

The company raised $300M in its Series A from Alibaba, Tencent, and others in October 2023, reaching a reported $1.8B valuation. It initially produced general-purpose open-weight LLMs (Baichuan-7B through Baichuan-4) optimized for Chinese and English, but made a distinctive strategic pivot into medical AI starting in 2024.

The Baichuan-M series (M1, M2, M3) represents this medical specialization, with Baichuan-M3 achieving leading performance on medical benchmarks. The company positions itself as building "AI for healthcare" rather than competing head-on in the crowded general-purpose LLM market. This pivot leverages Wang Xiaochuan's long-standing interest in AI-powered search and information retrieval applied to clinical settings.

Website Wikipedia HuggingFace GitHub
open-weightmedical-aimultilingual

People

Outputs (16)

Baichuan-M4

paper

"A Clinical-Grade Medical Agent System for Continuous Care" — the fourth generation of Baichuan's medical line. Introduces Baichuan-Harness (a u…

Baichuan-M3 2

paper model

235-billion-parameter open-source medical model that conducts full clinical dialogue with proactive information acquisition, long-horizon reasoning, a…

★ 245 ↓ 1.8k

Baichuan-M2 2

paper model

Medical-enhanced reasoning model built upon Qwen2.5-32B with an innovative Large Verifier System comprising a Patient Simulator and Clinical Rubrics G…

★ 212 ↓ 938 📄 1

BaichuanMed-OCR

model

Vision-language models fine-tuned from Qwen2.5-VL for medical report OCR. Trained on curated medical report datasets to accurately recognize complex t…

↓ 169

Baichuan-Audio 2

paper model

Open-source end-to-end audio large language model integrating speech understanding and generation for real-time bilingual Chinese-English dialogue. Us…

★ 222 ↓ 103

Baichuan-Omni-1.5 2

paper model

End-to-end trained omni-modal model supporting text, image, video, and audio input with text and audio output. Built upon Qwen2.5-7B, uses a custom au…

★ 191 ↓ 1.8k

Baichuan-M1 3

paper model

First open-source large language model developed from scratch for medical applications. Trained on 20 trillion tokens of both general and medical-spec…

↓ 821 📄 5

Baichuan4-Finance 2

paper model

Financial domain-specific model built upon Baichuan4-Turbo using continual pre-training with a novel domain self-constraint training strategy. Integra…

📄 1

Baichuan-Omni 2

paper model

First open-source 7B omni-modal large language model capable of concurrently processing and analyzing image, video, audio, and text modalities. Uses a…

★ 273

Baichuan 4

model

Fourth-generation proprietary base model from Baichuan Intelligence with significant improvements over the previous generation: general ability improv…

Baichuan 3

model

Large language model with over 100 billion parameters. Claimed to surpass GPT-4 in Chinese tasks on evaluations such as CMMLU, GAOKAO, and AGI-Eval.

Baichuan-NPC

model

Role-playing large model designed for personalized character customization in game and interactive scenarios. Provides flexible NPC dialogue capabilit…

Baichuan 2 3

paper model

Second generation of open-source large language models with 7B and 13B parameters, trained from scratch on 2.6 trillion tokens. Matches or outperforms…

★ 4.1k ↓ 210.8k 📄 125

Baichuan-53B

model

53-billion-parameter proprietary model, the third large model product from Baichuan Intelligence. Released alongside the launch of Baichuan's first ro…

Baichuan-13B

model

Open-source 13-billion-parameter language model supporting Chinese and English. Released under an Apache 2.0 license with both base and chat variants.…

★ 2.9k ↓ 14.7k

Baichuan-7B

model

Open-source 7-billion-parameter language model trained on 1.2 trillion tokens of Chinese and English data. Released under an Apache 2.0 license, it wa…

★ 5.7k ↓ 148.4k