百川智能发布 Baichuan 开源大模型

2023 年 6 月 15 日，百川智能发布了 Baichuan-7B，一个开源的大语言模型。这家公司的创始人王小川，是搜狗的联合创始人和前 CEO——一位互联网搜索老兵。他 2023 年 4 月宣布创业做 AI，目标直指「接近 ChatGPT 水平的中文大模型」，而第一个动作，就是开源。

在 2023 年春天的中国大模型市场，这是一个反直觉的选择。当时的主流叙事是「大模型能力竞赛」，各家公司比拼参数规模与对话效果，开源更多被视为 Meta 等海外公司的玩法。王小川却选择了另一条路：第一天就开源。Baichuan-7B 之后，百川迅速推出 13B 和 53B 版本，形成从 7B 到 53B 的完整梯度。

开源带来的是快速扩散。Baichuan 系列在中文开发者社区里迅速传播——研究者拿它微调，开发者拿它做垂直应用，企业拿它做私有化部署。相比必须走 API 的闭源模型，开源权重意味着数据不出门、二次开发自由。这个特性在重视数据安全的企业场景里尤其有价值。百川同时推进商业 API 与 ToB 服务，走「开源起盘、双轨变现」的路。

百川的打法验证了一个判断：在中国市场，开源大模型不是海外巨头的专利，本土公司同样能靠它建立生态。它的发布时间早于不少同行，让中文开源基座从几乎空白变成了一个快速成长的赛道。此后，通义千问、DeepSeek、GLM 等一批国产开源模型陆续发布，中国开源模型生态由此成形。

回看 Baichuan 的发布，它没有成为国产模型的性能巅峰，却示范了一种重要的入场策略：在能力暂时追不上最头部时，用开源换取生态、用生态积累数据与反馈。后来很多中国 AI 公司的成功路径里，都能看到这套打法的影子——先开源聚拢社区，再商业化变现。王小川带着搜索老兵的经验，给中国大模型市场补上了「开源」这块拼图。

On June 15, 2023 Baichuan released Baichuan-7B, an open-source large language model. The company's founder, Wang Xiaochuan, is the co-founder and former CEO of Sogou—an internet search veteran. He announced his AI startup in April 2023, aiming directly at "Chinese large models approaching ChatGPT level," and his first move was open source.

In China's spring 2023 large-model market, that was a counterintuitive choice. The dominant narrative was a capability race—companies competed on parameter scale and dialogue quality, and open source was seen more as an overseas play by Meta and others. Wang Xiaochuan chose another road: open from day one. After Baichuan-7B, Baichuan quickly shipped 13B and 53B versions, forming a complete gradient from 7B to 53B.

Open source brought rapid diffusion. The Baichuan series spread fast through Chinese developer communities—researchers fine-tuned it, developers built vertical applications, enterprises used it for private deployment. Compared to closed models that must go through APIs, open weights mean data stays in-house and secondary development is free. That property is especially valuable in data-security-sensitive enterprise scenarios. Baichuan simultaneously pushed commercial APIs and B2B services, walking an "open-source start, dual-track monetization" path.

Baichuan's play validated a judgment: in China, open-source large models are not an overseas exclusive; domestic companies can build ecosystems with them just as well. Its release came earlier than many peers, turning Chinese open bases from near-zero into a fast-growing track. Tongyi Qianwen, DeepSeek, GLM and others followed, and China's open-model ecosystem took shape.

Looking back at Baichuan's release, it did not become the performance peak of domestic models, but it demonstrated an important entry strategy: when capability cannot yet catch the very top, trade open source for ecosystem, and trade ecosystem for data and feedback. Many later Chinese AI success paths carry the shadow of this play—open source to gather community first, then commercialize. With a search veteran's experience, Wang Xiaochuan added the "open source" piece to China's large-model puzzle.