百度发布文心 4.5 与 X1

2025 年 3 月 16 日，文心一言的模型选择里多了两个名字。文心 4.5 面向原生多模态理解，文心 X1 强调“深度思考”；两者同时进入产品与百度智能云千帆，并向普通用户免费开放。没有等待一场大型发布会结束，模型先变成了可以点击的选项。

两个名字承担了不同的产品问题。4.5 要处理文字之外的图像等输入，把百度已有的搜索、内容与多模态能力合到一个旗舰档；X1 则回应推理模型热潮，让用户在更直接的回答与更长的推理过程之间选择。它们没有被包装成一个无所不能的单一模型。百度把“看得更多”和“想得更久”拆成两条路线，再放回同一个入口。

免费是这次上线最具体的分发动作。DeepSeek-R1 之后，用户已经开始用推理能力、响应速度和价格重新比较国内模型。百度拥有搜索流量、文心一言用户与千帆企业平台，可以用已有入口迅速把新模型送到大量账户，而不必等待开发者先更换应用。对普通用户，切换只发生在下拉菜单；对企业，模型标识、价格、延迟和输出风格都需要重新进入测试清单。

官方材料强调 4.5 的多模态能力和 X1 的推理性价比，这些仍是厂商自己的定位。免费试用能够扩大样本，却不能证明每一种任务都更可靠；推理过程更长也可能带来延迟与成本。真正的比较需要把同一批文档、图片、代码和业务问题交给两个模型，看它们各自在什么地方出错，而不是只沿用发布页的排名。

百度同时给出后续开放 4.5 系列的计划。这个承诺与当日可用的免费产品属于两种交付：一个已经出现在聊天与云平台里，另一个要等权重、许可证和运行文档真正发布后才算兑现。将二者写在同一策略中，说明百度也感受到开放权重生态的压力；但在时间线上，承诺仍应保持将来时。

文心 4.5 与 X1 的发布价值，最终落在一次大厂式的快速分发上。百度没有只用一份技术报告回应新的竞争，而是让多模态和推理在同一天成为现有用户可试的两个按钮。它证明入口仍是一种模型能力：实验室可以训练出新系统，平台公司则能决定多少人会在第二天真正用到它。

On 16 March 2025, two new names appeared in ERNIE Bot’s model selector. ERNIE 4.5 targeted native multimodal understanding; ERNIE X1 emphasized “deep thinking.” Both entered the consumer product and Baidu AI Cloud’s Qianfan platform, and both were made free for ordinary users. The models became clickable options without waiting for the end of a large keynote.

The two names carried different product questions. ERNIE 4.5 was meant to handle inputs beyond text, bringing images and other modalities into a flagship connected to Baidu’s search and content systems. X1 answered the reasoning-model wave by giving users a separate path for tasks that might benefit from a longer inference process. Baidu did not present one system as universally sufficient. It divided “see more” and “think longer” into two routes, then placed them behind the same doorway.

Free access was the release’s most concrete distribution move. After DeepSeek-R1, Chinese users had begun comparing models with a new combination of reasoning ability, response time, and price. Baidu already controlled search traffic, an ERNIE user base, and the Qianfan enterprise platform. It could move new models into many existing accounts without waiting for developers to adopt another application. For a consumer, the change was a dropdown selection. For an enterprise, model identifiers, price, latency, and output behavior all returned to the integration checklist.

Baidu positioned 4.5 around multimodal capability and X1 around reasoning cost-performance. Those remained publisher claims. Free trials could create a large testing population, but did not demonstrate equal reliability on every task. Longer reasoning could also increase latency and resource use. Meaningful comparison required submitting the same documents, images, code, and business questions to both models, then observing where each failed—not merely preserving the rankings on a launch page.

The company also announced a plan to open the 4.5 family later. That promise and the free product available that day were two different deliveries. One had appeared in chat and cloud interfaces; the other would count only when weights, licensing, and runnable documentation actually arrived. Placing both in one strategy showed how strongly open-weight ecosystems were influencing Baidu. On a timeline, however, the promise still belonged in the future tense.

The release of ERNIE 4.5 and X1 ultimately demonstrated a large platform’s ability to distribute quickly. Baidu did not answer the new competition with a technical report alone. It made multimodality and reasoning into two buttons that existing users could test on the same day. Access itself remained a form of model power: laboratories could train new systems, while platform companies could determine how many people would actually use them the following morning.