发布日期: 2026-01-09
版本号: v0.7.0

Open WebUI 发布了 0.7.0 版本，带来了一系列重要更新。主要新增功能包括支持使用内置工具进行多步骤任务的原生函数调用、从笔记和聊天记录中检索相关上下文、以及直接从知识库中搜索和检索文档。系统性能得到了显著提升，数据库连接处理、页面加载速度和查询优化均有大幅改进。管理员现在可以限制文件夹中的文件数量、更精细地控制用户权限（如访问控制、共享设置、功能开关）、配置各种外部服务参数（如Whisper、Firecrawl、Jina搜索）以及自动创建管理员账户。此外，还修复了大量问题，涵盖XSS安全漏洞、工具权限检查、聊天稳定性、文件处理、引用显示以及多种编辑和上传问题。该版本包含一个重要的后端数据库连接处理变更，要求多实例部署时必须同步更新所有节点。其他变更还包括文档分块方式的调整、管理页面路径的更正以及群组共享默认权限的改变。

更新内容 （中文）

[0.7.0] - 2026-01-09

新增功能

  • 🤖 内置工具的原生函数调用。 用户现在可以在单次对话中要求模型执行结合网页研究、知识库查询、笔记和图像生成的多步骤任务——例如，“研究X的最新进展，将关键发现保存到笔记，并生成信息图”。需要支持原生函数调用的模型，并在聊天控件中将函数调用模式设置为“原生”。#19397，提交
  • 🧠 用户现在可以要求模型从他们的笔记、过去的聊天和频道消息中查找相关上下文——例如，“我上周关于项目X讨论了什么？”或“找到我为Y brainstorming想法的对话”。提交
  • 📚 用户现在可以要求模型搜索他们的知识库并检索文档，而无需手动附加文件——例如，“在我们的API文档中找到关于认证的部分”或“我们的内部指南对X有什么规定？”提交
  • 💭 支持交错思考的模型的用户现在能从多步骤工作流中获得更精细的结果，因为模型可以在决定下一步操作之前分析每个工具的输出。
  • 🔍 当模型调用网页搜索时，搜索结果会以可点击的引用形式实时显示，以进行完整的来源验证。提交
  • 🎚️ 用户现在可以通过模型编辑器的功能设置，为每个模型选择性地禁用特定内置工具（时间戳、记忆、聊天历史、笔记、网页搜索、知识库）。提交
  • 👁️ 待处理的工具调用现在会在响应生成期间显示，以便用户知道哪些工具正在被调用。提交
  • 📁 管理员现在可以使用“FOLDER_MAX_FILE_COUNT”设置来限制上传到文件夹的文件数量，防止批量上传导致资源耗尽。#19810，提交，提交
  • ⚡ 通过彻底重新设计的数据库连接处理，用户在整个应用程序中体验到变革性的速度提升，带来明显更快的页面加载、丝滑流畅的操作体验，以及在用户管理和批量数据处理等密集型操作期间坚如磐石的稳定性。提交，提交，提交，提交，提交，提交
  • 🚀 通过动态加载文档处理库，减少初始包大小，用户显著体验到更快的初始页面加载时间。#20200，#20202，#20203，#20204
  • 💨 通过优化的数据库查询（消除了N+1查询模式），管理员体验到显著更快的用户列表加载速度，无论用户数量多少，查询次数都从1+N减少到仅2次总计查询。#20427
  • 📋 笔记现在加载更快，通过优化的数据库查询批量处理用户查找，而不是单独获取每个笔记的作者。提交
  • 💬 频道消息、置顶消息和线程回复现在加载更快，通过批量处理用户查找，而不是对每条消息进行单独查询。#20458，#20459，#20460
  • 🔗 用户现在可以点击引用内容链接直接跳转到源文档的相关部分，并自动高亮文本，使得对照原始来源验证AI回复变得更加容易。#20116，提交
  • 📌 用户现在可以直接从工作空间模型页面和管理设置模型页面固定或隐藏模型，使得管理哪些模型出现在侧边栏中更加容易，无需切换到聊天界面。#20176
  • 🔎 管理员现在可以使用管理设置侧边栏中的新搜索栏快速查找设置，该栏支持按类别名称和相关关键词（如音频的“whisper”或文档的“rag”）进行模糊过滤。#20434
  • 🎛️ 用户现在可以在工作空间模型列表中查看只读模型，并有明确的“只读”徽章指示编辑受限。#20243，#20369
  • 📝 用户现在可以在工作空间提示词列表中查看只读提示词，并有明确的“只读”徽章指示编辑受限。#20368
  • 🔧 用户现在可以在工作空间工具列表中查看只读工具，并有明确的“只读”徽章指示编辑受限。#20243，#20370
  • 📂 文件搜索现在显著加快，尤其对于拥有大量文件集合的用户。提交
  • 🏆 评估排行榜现在在后端计算埃洛评级，而不是在浏览器中，提高了性能，并支持通过语义搜索进行基于主题的模型排名。#15392，#20476，提交
  • 📊 评估排行榜现在包含每个模型的活动图表，以发散条形图显示每日胜负，并提供30天、1年和所有时间视图，较长的时间范围使用每周聚合。
  • 🎞️ 用户现在可以上传动画GIF和WebP格式作为模型个人资料图片，通过跳过这些文件类型的缩放处理来保留动画。提交
  • 📸 用户上传用户、模型和竞技场模型的个人资料图片时，现在受益于80%质量的WebP压缩（而不是JPEG），从而显著减小文件大小并加快上传速度，同时保持视觉质量。提交
  • ⭐ 动作函数开发者现在可以使用新的“chat:message:favorite”事件更新消息收藏状态，从而开发固定/取消固定消息动作，避免了前端自动保存带来的竞态条件。#20375
  • 🌐 使用具有网页搜索能力的OpenAI兼容模型的用户现在会在界面中看到以来源形式显示的URL引用。#20172，提交
  • 📰 用户现在可以使用X按钮永久关闭“新功能”变更日志模态框，其行为与“好的，我们开始吧！”按钮匹配。#20258
  • 📧 管理员现在可以直接从管理面板配置账户待定覆盖层中显示的管理员联系电子邮件，而不再仅通过环境变量。#12500，#20260
  • 📄 管理员现在可以通过新的“ENABLE_MARKDOWN_HEADER_TEXT_SPLITTER”设置启用Markdown标题文本分割作为预处理步骤，与字符或标记分割结合使用。提交，提交，提交
  • 🧩 管理员现在可以使用“CHUNK_MIN_SIZE_TARGET”设置设置最小块大小目标，将小的Markdown标题块与相邻块合并，从而通过消除微小无意义的片段来提高检索质量，显著加快文档处理和嵌入速度，降低存储成本，并降低嵌入API成本或本地计算要求。#19595，#20314，提交
  • 💨 管理员现在可以通过将“RAG_SYSTEM_CONTEXT”设置为true来启用KV前缀缓存优化，这会将RAG上下文注入系统消息（而不是用户消息），使模型能够重用缓存的标记来回答后续问题，而不是在每一轮次都重新处理整个上下文，从而显著提高响应速度并降低基于云的模型成本。#20301，#20317
  • 🖼️ 管理员和动作开发者现在可以使用步骤参数按请求控制图像生成的去噪步数，允许动作和API调用覆盖ComfyUI和Automatic1111引擎的全局IMAGE_STEPS配置。#20337
  • 🗄️ 运行多Pod部署的管理员现在可以使用“ENABLE_DB_MIGRATIONS”环境变量指定一个主Pod来处理数据库迁移。提交
  • 🎙️ 管理员现在可以使用“WHISPER_COMPUTE_TYPE”环境变量配置Whisper的计算类型，以修复与CUDA/GPU部署的兼容性问题。提交
  • 🔍 管理员现在可以使用“SENTENCE_TRANSFORMERS_CROSS_ENCODER_SIGMOID_ACTIVATION_FUNCTION”环境变量控制CrossEncoder重排模型的Sigmoid归一化，默认为启用状态，以确保MS MARCO模型的适当相关性阈值行为。#19999 #20228
  • 🔒 管理员现在可以使用“REQUESTS_VERIFY”环境变量禁用外部工具的SSL证书验证，从而启用与Tika、Ollama嵌入和外部重排器的自签名证书集成。#19968，提交
  • 📈 管理员现在可以使用“ENABLE_AUDIT_STDOUT”和“ENABLE_AUDIT_LOGS_FILE”环境变量控制审计日志输出目标，允许审计日志发送到容器日志以实现集中式日志记录系统。#20114，提交
  • 🛡️ 管理员现在可以通过每个用户或每个组的权限限制非管理员用户访问界面设置。#20424
  • 🧠 管理员现在可以全局启用或禁用记忆功能，并通过每个用户或每个组的权限控制访问，禁用功能时个性化选项卡会自动隐藏。#20462
  • 🟢 管理员现在可以通过管理设置中的“ENABLE_USER_STATUS”设置全局启用或禁用用户状态可见性。#20488
  • 🪝 频道管理员现在可以创建网络钩子，允许外部服务在无需认证的情况下向频道发送消息。提交
  • 📄 在模型编辑器中，用户现在可以禁用“文件上下文”功能，以跳过自动文件内容提取和注入，改为转发带有文件附件元数据的原始消息，供自定义工具或未来的内置文件访问工具使用。提交，文档提交
  • 🔊 在模型编辑器中，用户现在可以为每个模型配置特定的TTS语音，覆盖用户偏好和全局默认值，为不同AI角色提供独特的声音。#3097，提交
  • 👥 管理员现在拥有三个精细的组共享权限选项（而不是简单的开/关切换），允许他们为每个组选择“无人”、“成员”或“任何人”来决定谁可以共享内容到该组。提交
  • 📦 管理员现在可以将知识库导出为包含文本文件的zip文件，用于备份和归档。#20120，提交
  • 🚀 管理员现在可以通过“WEBUI_ADMIN_EMAIL”、“WEBUI_ADMIN_PASSWORD”和“WEBUI_ADMIN_NAME”环境变量在启动时自动创建管理员账户，从而实现无需暴露注册页面的无头和自动化部署。#17654，提交
  • 🦆 管理员现在可以为DDGS选择特定的搜索后端（而不是随机选择），选项包括Bing、Brave、DuckDuckGo、Google、Wikipedia、Yahoo等。#20330，#20366
  • 🧭 管理员现在可以使用“JINA_API_BASE_URL”环境变量配置自定义Jina搜索API端点，从而支持特定区域的部署，例如欧盟数据处理。#19718，提交
  • 🔥 管理员现在可以使用“FIRECRAWL_TIMEOUT”环境变量配置Firecrawl超时值，以控制网页抓取等待时间。#19973，提交
  • 💾 管理员现在可以使用openGauss作为知识库文档存储和检索的向量数据库后端。#20179
  • 🔄 在整个应用程序中实施了各种改进，以增强性能、稳定性和安全性。
  • 📊 用户现在可以将他们的匿名使用统计同步到Open WebUI社区平台，以支持社区排行榜、推动模型评估，并为造福每个人的开源AI研究做出贡献，同时保持对话完全私密（仅共享模型名称、消息计数和评级等元数据）。通过分享您的统计数据，您正在帮助社区识别哪些模型表现最好，为透明的AI基准测试做出贡献，并支持让AI为所有人变得更好的集体努力。您还可以将您的统计数据下载为JSON格式进行个人分析。
  • 🌐 德语、葡萄牙语（巴西）、西班牙语、简体中文、繁体中文和波兰语的翻译得到了增强和扩展。

修复问题

  • 🔊 文本转语音现在除了标点符号外，还会在换行符处分割，因此markdown项目符号和列表会被作为单独的句子朗读，而不是合并在一起。#5924，提交
  • 🔒 用户现在受到iFrame嵌入（用于引用和响应消息）中存储型XSS漏洞的保护，通过可配置的同源沙箱设置（而不是硬编码值）实现。#20209，#20210
  • 🔑 图像生成、网页搜索和音频（TTS/STT）API端点现在在后端强制执行权限检查，关闭了一个安全漏洞，即之前禁用的功能可以通过直接API调用访问。#20471
  • 🛠️ 工具和工具服务器（MCP和OpenAPI）现在在后端强制执行访问控制检查，确保用户只能访问他们有权使用的工具，即使通过直接API调用也无法例外。#20443，提交
  • 🔁 使用原生函数调用时，系统提示不再重复，修复了在工具调用工作流期间提示被应用两次的问题。提交
  • 🗂️ 当未设置“FOLDER_MAX_FILE_COUNT”时，知识库上传到文件夹不再失败，修复了默认空值导致所有上传出错的问题。提交
  • 📝 聊天输入中的“创建笔记”按钮现在对于没有笔记权限的用户正确隐藏，而不是显示并在点击时返回401错误。#20486，提交
  • 📊 评估页面现在不再因管理员拥有大量反馈数据而崩溃，因为排行榜现在只获取所需的最小字段，而不是加载整个对话快照。#20476，#20489，提交
  • 💬 用户现在可以再次正确导出聊天、使用提问/解释弹窗和查看聊天列表，因为这些功能因最近的重构更改导致500和400服务器错误而失效。#20146，#20205，#20206，#20212
  • 💭 用户在背景操作（如图像生成）期间切换聊天时不再遇到数据损坏，即来自一个聊天的消息会出现在另一个聊天的历史记录中。#20266
  • 🛡️ 用户不再遇到严重的聊天稳定性错误，包括循环消息依赖导致的重复键错误、聊天加载期间的空消息访问错误以及聊天概览可视化中的错误。#20268
  • 📡 使用频道的用户在获取线程回复时不再经历无限递归和连接池耗尽，防止了聊天历史记录加载期间的RecursionError崩溃。#20299，提交
  • 📎 用户在查看具有未定义URL属性的文件附件消息时不再遇到TypeError崩溃。#20343
  • 🔐 使用MCP集成的用户现在在访问令牌过期后能通过适当的受保护资源发现体验到可靠的OAuth 2.1令牌刷新，防止了导致会话被删除的集成失败。#19794，#20138，#20291，提交，提交
  • 📚 属于多个组的用户现在可以看到与这些组共享的知识库，修复了当知识库与多个组共享时会消失的问题。#20124，#20229，提交
  • 📂 用户现在在聊天输入中悬停在#文件引用上时能看到正确的知识库名称，而不是“undefined”。#20329，#20333
  • 📋 用户现在在笔记的时间范围分组内看到正确按时间顺序排列的笔记，修复了插入顺序未被保留的问题。提交
  • 📑 协作笔记的用户现在在从HTML和JSON格式初始化时能体验到正确的内容同步，修复了协作编辑会话中的同步失败。提交
  • 🔎 搜索笔记的用户现在能找到连字符单词和带空格的变体，因此搜索“todo”现在能找到“to-do”和“to do”。提交
  • 📥 用户在初始处理失败后重新上传文件时不再经历误报的重复文件警告，因为现在只在处理成功完成后才存储文件哈希。#19264，#20282，提交
  • 💾 用户体验到显著的页面加载性能提升，因为模型个人资料图片现在能在浏览器中正确缓存，避免不必要的图片重新获取。提交
  • 🎨 用户现在可以成功编辑上传的图像（而不是生成新图像），修复了文件存储重构引入的一个问题：类型为“file”且content_type以“image/”开头的图像未被识别为可编辑图像。#20237，#20169，#20239，提交
  • 🌐 使用波斯语和阿拉伯语撰写的用户现在在笔记部分能看到正确显示的从右到左文本，通过自动文本方向检测实现。#19743，#20102，提交
  • 🤖 用户现在可以在频道中成功@提及模型，而不是经历静默失败。提交
  • 📋 Windows用户现在使用{{CLIPBOARD}}变量时能看到正确保留的换行符，通过CRLF到LF的规范化实现。#19370，#20283
  • 📁 用户现在能看到知识选择器下拉菜单在创建文件夹模态框上方正确显示，而不是被其遮挡。#20219，#20213
  • 🌅 非PNG格式（如SVG、JPEG和GIF）的个人资料图片现在能正确显示，而不是出现损坏。#20171
  • 🆕 具有禁用临时聊天权限的非管理员用户现在可以成功创建新聊天并使用侧边栏的固定模型。#20336，#20367，提交
  • 🎛️ 用户现在可以在聊天中成功使用工作空间模型，修复了使用自定义模型预设时出现的“模型未找到”错误。#20340，#20344，提交，提交
  • 🔁 当父消息在聊天历史中丢失或损坏时，用户现在可以重新生成消息而不会崩溃。#20264
  • ✏️ 用户不再因插件或异步方法在编辑器被销毁后尝试访问编辑器而经历TipTap富文本编辑器崩溃并出现“editor view is not available”错误。#20266
  • 📗 启用了绕过访问控制的管理员现在正确拥有对所有知识库的写权限。#20371
  • 🔍 使用本地CrossEncoder重排模型的管理员现在通过MS MARCO模型分数归一化到0-1范围（通过Sigmoid激活）看到适当的相关性阈值行为。#19999，#20228
  • 🎯 使用本地SentenceTransformers嵌入引擎的管理员现在受益于适当的批量大小设置，防止了默认批量大小32导致的过度内存使用。#20053，#20054，提交
  • 🔧 在离线模式或受限环境（如uv、poetry和NixOS）中的管理员和用户，当工具和函数有前置元要求时不再经历崩溃，因为现在在启用离线模式时会跳过pip安装。#20320，#20321，提交
  • 📄 管理员现在可以正确配置MinerU文档解析服务，因为MinerU Cloud API密钥字段现在在管理面板文档设置中可用。#20319，#20328
  • ⚠️ 管理员不再看到密码验证正则表达式模式中无效转义序列的语法警告。#20298，提交
  • 🎨 使用ComfyUI工作流的用户现在在聊天中只看到预期的最终输出图像，而不是来自中间处理节点（如蒙版、裁剪或分割预览）的重复图像。#20158，#20182
  • 🖼️ 启用图像生成的用户不再看到误报的视觉能力警告，允许他们在生成图像后发送后续消息，并将图像发送到非视觉模型进行图像编辑。#20129，#20256
  • 🔌 当外部工具使用无效或MCP风格配置与OpenAPI连接类型时，管理员不再经历无限加载屏幕。#20207，#20257
  • 📥 管理员通过URL上传GGUF模型时，在SHA256验证期间不再遇到TypeError崩溃，修复了500内部服务器错误崩溃。#20263
  • 🚦 使用Brave搜索的用户在遇到速率限制时现在会自动重试（延迟1秒），防止连续请求超过每秒1次请求限制时的失败，尽管这仅在网页搜索并发设置为最大1时可靠工作。#15134，#20255
  • 🗄️ 使用Redis Sentinel部署的管理员在WebSocket断开连接期间不再因YDocManager中不正确的异步生成器处理而经历崩溃。#20142，#20145
  • 🔐 使用SCIM组管理的管理员在处理没有成员的组时不再遇到500错误。#20187
  • 🔗 用户现在能体验到来自AI模型的更可靠的引用，尤其是在使用较小或较弱模型时，这些模型可能无法完美格式化引用文献。提交
  • 🕸️ 管理员现在可以成功保存WebSearch设置，而不会遇到域名过滤列表、YouTube语言设置或超时值的验证错误。#20422
  • 📦 使用uv包管理器安装的管理员在移除导致冲突的已弃用依赖项后，现在能成功安装。#20177，#20192
  • ⏱️ 使用自定义“AIOHTTP_CLIENT_TIMEOUT”设置的管理员现在看到配置的超时正确应用于嵌入生成、OAuth发现、网络钩子调用和工具/函数加载，而不是回退到默认的300秒超时。提交

变更内容

  • ⚠️ 此版本包含后端数据库连接处理的重大检修，要求多工作者、多服务器或负载平衡部署中的所有实例同时更新；运行混合版本将因新旧实例之间不兼容的数据库连接管理而导致失败。
  • 📝 之前使用独立“Markdown (Header)”文本分割器的管理员现在必须切换到“字符”或“标记”模式，并启用新的“ENABLE_MARKDOWN_HEADER_TEXT_SPLITTER”切换，因为文档分块现在将Markdown标题分割作为字符或标记分割之前的预处理步骤。提交，提交，提交
  • 🖼️ 用户在聊天消息界面中不再看到“生成图像”操作按钮；应使用自定义函数。提交
  • 🔗 管理员将在新URL“/admin/evaluations/feedback”找到管理评估页面（而不是“/admin/evaluations/feedbacks”），以使用该单词正确的不可数形式。#20296
  • 🔐 当图像生成、网页搜索或音频功能在管理UI中被禁用时，直接调用这些功能的API的脚本或集成现在将收到403禁止访问错误，因为已添加后端权限执行以匹配前端限制。#20471
  • 👥 默认组共享权限从“成员”更改为“任何人”，这意味着用户现在可以将内容共享到任何配置了“任何人”权限的组，无论他们是否属于该组。提交

[!TIP]
寻找企业计划？ – 今天就联系我们的销售团队！

获取增强功能，包括自定义主题和品牌、服务水平协议（SLA）支持等等！

👏 巨大的感谢我们了不起的贡献者们的辛勤工作和奉献，使这次发布成为可能： @ShirasawaSama, @BLACKTHOMAS, @imsamurai, @jannikstdl, @silentoplayz, @joaoback, @rgaricano, @silentoplayz, @Classic298, @majiayu000, @Ithanil, @vincentsunx, @landerrosette, @BarerRocket2678, @rgaricano

赞助商 🙌

🚀 我们要向我们了不起的赞助商们的慷慨支持表示衷心的感谢（注意：我们已从本列表中排除了私人赞助商。如果您希望在此处展示，请随时联系我们！）

@amd, @roosi-gmbh, @rndmcnlly, @feddersen-group, @SRKConsulting, @mushmax, @icanbwell, @loitragg, @haka4githu6, @silenceroom, @crizCraig, @GenieDev101, @CubicleGangster, @TOWaB, @Classic298, @comet-ml, @FominVO, @catalyst-net-nz, @JrdnHans, @hkosm, @VeroFess, @joaoback, @walsermedia, @LuxTronic

更新内容 （原始）

[0.7.0] - 2026-01-09

Added

  • 🤖 Native Function Calling with Built-in Tools. Users can now ask models to perform multi-step tasks that combine web research, knowledge base queries, note-taking, and image generation in a single conversation—for example, “research the latest on X, save key findings to a note, and generate an infographic.” Requires models with native function calling support and function calling mode set to “Native” in Chat Controls. #19397, Commit
  • 🧠 Users can now ask the model to find relevant context from their notes, past chats, and channel messages—for example, “what did I discuss about project X last week?” or “find the conversation where I brainstormed ideas for Y.” Commit
  • 📚 Users can now ask the model to search their knowledge bases and retrieve documents without manually attaching files—for example, “find the section about authentication in our API docs” or “what do our internal guidelines say about X?” Commit
  • 💭 Users with models that support interleaved thinking now get more refined results from multi-step workflows, as the model can analyze each tool’s output before deciding what to do next.
  • 🔍 When models invoke web search, search results appear as clickable citations in real-time for full source verification. Commit
  • 🎚️ Users can selectively disable specific built-in tools (timestamps, memory, chat history, notes, web search, knowledge bases) per model via the model editor’s capabilities settings. Commit
  • 👁️ Pending tool calls are now displayed during response generation, so users know which tools are being invoked. Commit
  • 📁 Administrators can now limit the number of files that can be uploaded to folders using the “FOLDER_MAX_FILE_COUNT” setting, preventing resource exhaustion from bulk uploads. #19810, Commit, Commit
  • ⚡ Users experience transformative speed improvements across the entire application through completely reengineered database connection handling, delivering noticeably faster page loads, butter-smooth interactions, and rock-solid stability during intensive operations like user management and bulk data processing. Commit, Commit, Commit, Commit, Commit, Commit
  • 🚀 Users experience significantly faster initial page load times through dynamic loading of document processing libraries, reducing the initial bundle size. #20200, #20202, #20203, #20204
  • 💨 Administrators experience dramatically faster user list loading through optimized database queries that eliminate N+1 query patterns, reducing query count from 1+N to just 2 total queries regardless of user count. #20427
  • 📋 Notes now load faster through optimized database queries that batch user lookups instead of fetching each note’s author individually. Commit
  • 💬 Channel messages, pinned messages, and thread replies now load faster through batched user lookups instead of individual queries per message. #20458, #20459, #20460
  • 🔗 Users can now click citation content links to jump directly to the relevant portion of source documents with automatic text highlighting, making it easier to verify AI responses against their original sources. #20116, Commit
  • 📌 Users can now pin or hide models directly from the Workspace Models page and Admin Settings Models page, making it easier to manage which models appear in the sidebar without switching to the chat interface. #20176
  • 🔎 Administrators can now quickly find settings using the new search bar in the Admin Settings sidebar, which supports fuzzy filtering by category names and related keywords like “whisper” for Audio or “rag” for Documents. #20434
  • 🎛️ Users can now view read-only models in the workspace models list, with clear “Read Only” badges indicating when editing is restricted. #20243, #20369
  • 📝 Users can now view read-only prompts in the workspace prompts list, with clear “Read Only” badges indicating when editing is restricted. #20368
  • 🔧 Users can now view read-only tools in the workspace tools list, with clear “Read Only” badges indicating when editing is restricted. #20243, #20370
  • 📂 Searching for files is now significantly faster, especially for users with large file collections. Commit
  • 🏆 The Evaluations leaderboard now calculates Elo ratings on the backend instead of in the browser, improving performance and enabling topic-based model ranking through semantic search. #15392, #20476, Commit
  • 📊 The Evaluations leaderboard now includes a per-model activity chart displaying daily wins and losses as a diverging bar chart, with 30-day, 1-year, and all-time views using weekly aggregation for longer timeframes.
  • 🎞️ Users can now upload animated GIF and WebP formats as model profile images, with animation preserved by skipping resize processing for these file types. Commit
  • 📸 Users uploading profile images for users, models, and arena models now benefit from WebP compression at 80% quality instead of JPEG, resulting in significantly smaller file sizes and faster uploads while maintaining visual quality. Commit
  • ⭐ Action Function developers can now update message favorite status using the new “chat:message:favorite” event, enabling the development of pin/unpin message actions without race conditions from frontend auto-save. #20375
  • 🌐 Users with OpenAI-compatible models that have web search capabilities now see URL citations displayed as sources in the interface. #20172, Commit
  • 📰 Users can now dismiss the “What’s New” changelog modal permanently using the X button, matching the behavior of the “Okay, Let’s Go!” button. #20258
  • 📧 Administrators can now configure the admin contact email displayed in the Account Pending overlay directly from the Admin Panel instead of only through environment variables. #12500, #20260
  • 📄 Administrators can now enable markdown header text splitting as a preprocessing step that works with either character or token splitting, through the new “ENABLE_MARKDOWN_HEADER_TEXT_SPLITTER” setting. Commit, Commit, Commit
  • 🧩 Administrators can now set a minimum chunk size target using the “CHUNK_MIN_SIZE_TARGET” setting to merge small markdown header chunks with neighbors, which improves retrieval quality by eliminating tiny meaningless fragments, significantly speeds up document processing and embedding, reduces storage costs, and lowers embedding API costs or local compute requirements. #19595, #20314, Commit
  • 💨 Administrators can now enable KV prefix caching optimization by setting “RAG_SYSTEM_CONTEXT” to true, which injects RAG context into the system message instead of user messages, enabling models to reuse cached tokens for follow-up questions instead of reprocessing the entire context on each turn, significantly improving response times and reducing costs for cloud-based models. #20301, #20317
  • 🖼️ Administrators and Action developers can now control image generation denoising steps per-request using a steps parameter, allowing Actions and API calls to override the global IMAGE_STEPS configuration for both ComfyUI and Automatic1111 engines. #20337
  • 🗄️ Administrators running multi-pod deployments can now designate a master pod to handle database migrations using the “ENABLE_DB_MIGRATIONS” environment variable. Commit
  • 🎙️ Administrators can now configure Whisper’s compute type using the “WHISPER_COMPUTE_TYPE” environment variable to fix compatibility issues with CUDA/GPU deployments. Commit
  • 🔍 Administrators can now control sigmoid normalization for CrossEncoder reranking models using the “SENTENCE_TRANSFORMERS_CROSS_ENCODER_SIGMOID_ACTIVATION_FUNCTION” environment variable, enabled by default for proper relevance threshold behavior with MS MARCO models. #19999 #20228
  • 🔒 Administrators can now disable SSL certificate verification for external tools using the “REQUESTS_VERIFY” environment variable, enabling integration with self-signed certificates for Tika, Ollama embeddings, and external rerankers. #19968, Commit
  • 📈 Administrators can now control audit log output destinations using “ENABLE_AUDIT_STDOUT” and “ENABLE_AUDIT_LOGS_FILE” environment variables, allowing audit logs to be sent to container logs for centralized logging systems. #20114, Commit
  • 🛡️ Administrators can now restrict non-admin user access to Interface Settings through per-user or per-group permissions. #20424
  • 🧠 Administrators can now globally enable or disable the Memories feature and control access through per-user or per-group permissions, with the Personalization tab automatically hidden when the feature is disabled. #20462
  • 🟢 Administrators can now globally enable or disable user status visibility through the “ENABLE_USER_STATUS” setting in Admin Settings. #20488
  • 🪝 Channel managers can now create webhooks to allow external services to post messages to channels without authentication. Commit
  • 📄 In the model editor users can now disable the “File Context” capability to skip automatic file content extraction and injection, forwarding raw messages with file attachment metadata instead for use with custom tools or future built-in file access tools. Commit, Docs:Commit
  • 🔊 In the model editor users can now configure a specific TTS voice for each model, overriding user preferences and global defaults to give different AI personas distinct voices. #3097, Commit
  • 👥 Administrators now have three granular group sharing permission options instead of a simple on/off toggle, allowing them to choose between “No one”, “Members”, or “Anyone” for who can share content to each group. Commit
  • 📦 Administrators can now export knowledge bases as zip files containing text files for backup and archival purposes. #20120, Commit
  • 🚀 Administrators can now create an admin account automatically at startup via “WEBUI_ADMIN_EMAIL”, “WEBUI_ADMIN_PASSWORD”, and “WEBUI_ADMIN_NAME” environment variables, enabling headless and automated deployments without exposing the signup page. #17654, Commit
  • 🦆 Administrators can now select a specific search backend for DDGS instead of random selection, with options including Bing, Brave, DuckDuckGo, Google, Wikipedia, Yahoo, and others. #20330, #20366
  • 🧭 Administrators can now configure custom Jina Search API endpoints using the “JINA_API_BASE_URL” environment variable, enabling region-specific deployments such as EU data processing. #19718, Commit
  • 🔥 Administrators can now configure Firecrawl timeout values using the “FIRECRAWL_TIMEOUT” environment variable to control web scraping wait times. #19973, Commit
  • 💾 Administrators can now use openGauss as the vector database backend for knowledge base document storage and retrieval. #20179
  • 🔄 Various improvements were implemented across the application to enhance performance, stability, and security.
  • 📊 Users can now sync their anonymous usage statistics to the Open WebUI Community platform to power community leaderboards, drive model evaluations, and contribute to open-source AI research that benefits everyone, all while keeping conversations completely private (only metadata like model names, message counts, and ratings are shared). By sharing your stats, you’re helping the community identify which models perform best, contributing to transparent AI benchmarking, and supporting the collective effort to make AI better for all. You can also download your stats as JSON for personal analysis.
  • 🌐 Translations for German, Portuguese (Brazil), Spanish, Simplified Chinese, Traditional Chinese, and Polish were enhanced and expanded.

Fixed

  • 🔊 Text-to-speech now correctly splits on newlines in addition to punctuation, so markdown bullet points and lists are spoken as separate sentences instead of being merged together. #5924, Commit
  • 🔒 Users are now protected from stored XSS vulnerabilities in iFrame embeds for citations and response messages through configurable same-origin sandbox settings instead of hardcoded values. #20209, #20210
  • 🔑 Image Generation, Web Search, and Audio (TTS/STT) API endpoints now enforce permission checks on the backend, closing a security gap where disabled features could previously be accessed via direct API calls. #20471
  • 🛠️ Tools and Tool Servers (MCP and OpenAPI) now enforce access control checks on the backend, ensuring users can only access tools they have permission to use even via direct API calls. #20443, Commit
  • 🔁 System prompts are no longer duplicated when using native function calling, fixing an issue where the prompt would be applied twice during tool-calling workflows. Commit
  • 🗂️ Knowledge base uploads to folders no longer fail when “FOLDER_MAX_FILE_COUNT” is unset, fixing an issue where the default null value caused all uploads to error. Commit
  • 📝 The “Create Note” button in the chat input now correctly hides for users without Notes permissions instead of showing and returning a 401 error when clicked. #20486, Commit
  • 📊 The Evaluations page no longer crashes when administrators have large amounts of feedback data, as the leaderboard now fetches only the minimal required fields instead of loading entire conversation snapshots. #20476, #20489, Commit
  • 💬 Users can now export chats, use the Ask/Explain popup, and view chat lists correctly again after these features were broken by recent refactoring changes that caused 500 and 400 server errors. #20146, #20205, #20206, #20212
  • 💭 Users no longer experience data corruption when switching between chats during background operations like image generation, where messages from one chat would appear in another chat’s history. #20266
  • 🛡️ Users no longer encounter critical chat stability errors, including duplicate key errors from circular message dependencies, null message access during chat loading, and errors in the chat overview visualization. #20268
  • 📡 Users with Channels no longer experience infinite recursion and connection pool exhaustion when fetching threaded replies, preventing RecursionError crashes during chat history loading. #20299, Commit
  • 📎 Users no longer encounter TypeError crashes when viewing messages with file attachments that have undefined URL properties. #20343
  • 🔐 Users with MCP integrations now experience reliable OAuth 2.1 token refresh after access token expiration through proper Protected Resource discovery, preventing integration failures that caused sessions to be deleted. #19794, #20138, #20291, Commit, Commit
  • 📚 Users who belong to multiple groups can now see Knowledge Bases shared with those groups, fixing an issue where they would disappear when shared with more than one group. #20124, #20229, Commit
  • 📂 Users now see the correct Knowledge Base name when hovering over # file references in chat input instead of “undefined”. #20329, #20333
  • 📋 Users now see notes displayed in correct chronological order within their time range groupings, fixing an issue where insertion order was not preserved. Commit
  • 📑 Users collaborating on notes now experience proper content sync when initializing from both HTML and JSON formats, fixing sync failures in collaborative editing sessions. Commit
  • 🔎 Users searching notes can now find hyphenated words and variations with spaces, so searching “todo” now finds “to-do” and “to do”. Commit
  • 📥 Users no longer experience false duplicate file warnings when reuploading files after initial processing failed, as the file hash is now only stored after successful processing completion. #19264, #20282, Commit
  • 💾 Users experience significantly improved page load performance as model profile images now cache properly in browsers, avoiding unnecessary image refetches. Commit
  • 🎨 Users can now successfully edit uploaded images instead of having new images generated, fixing an issue introduced by the file storage refactor where images with type “file” and content_type starting with “image/” weren’t being recognized as editable images. #20237, #20169, #20239, Commit
  • 🌐 Users writing in Persian and Arabic now see properly displayed right-to-left text in the notes section through automatic text direction detection. #19743, #20102, Commit
  • 🤖 Users can now successfully @ mention models in Channels instead of experiencing silent failures. Commit
  • 📋 Users on Windows now see correctly preserved line breaks when using the {{CLIPBOARD}} variable through CRLF to LF normalization. #19370, #20283
  • 📁 Users now see the Knowledge Selector dropdown correctly displayed above the Create Folder modal instead of being hidden behind it. #20219, #20213
  • 🌅 Users now see profile images in non-PNG formats like SVG, JPEG, and GIF displayed correctly instead of appearing broken. #20171
  • 🆕 Non-admin users with disabled temporary chat permissions can now successfully create new chats and use pinned models from the sidebar. #20336, #20367, Commit
  • 🎛️ Users can now successfully use workspace models in chat, fixing “Model not found” errors that occurred when using custom model presets. #20340, #20344, Commit, Commit
  • 🔁 Users can now regenerate messages without crashes when the parent message is missing or corrupted in the chat history. #20264
  • ✏️ Users no longer experience TipTap rich text editor crashes with “editor view is not available” errors when plugins or async methods try to access the editor after it has been destroyed. #20266
  • 📗 Administrators with bypass access control enabled now correctly have write access to all knowledge bases. #20371
  • 🔍 Administrators using local CrossEncoder reranking models now see proper relevance threshold behavior through MS MARCO model score normalization to the 0-1 range via sigmoid activation. #19999, #20228
  • 🎯 Administrators using local SentenceTransformers embedding engine now benefit from proper batch size settings, preventing excessive memory usage from the default batch size of 32. #20053, #20054, Commit
  • 🔧 Administrators and users in offline mode or restricted environments like uv, poetry, and NixOS no longer experience crashes when Tools and Functions have frontmatter requirements, as pip installation is now skipped when offline mode is enabled. #20320, #20321, Commit
  • 📄 Administrators can now properly configure the MinerU document parsing service as the MinerU Cloud API key field is now available in the Admin Panel Documents settings. #20319, #20328
  • ⚠️ Administrators no longer see SyntaxWarnings for invalid escape sequences in password validation regex patterns. #20298, Commit
  • 🎨 Users with ComfyUI workflows now see only the intended final output images in chat instead of duplicate images from intermediate processing nodes like masks, crops, or segmentation previews. #20158, #20182
  • 🖼️ Users with image generation enabled no longer see false vision capability warnings, allowing them to send follow-up messages after generating images and to send images to non-vision models for image editing. #20129, #20256
  • 🔌 Administrators no longer experience infinite loading screens when invalid or MCP-style configurations are used with OpenAPI connection types for external tools. #20207, #20257
  • 📥 Administrators no longer encounter TypeError crashes during SHA256 verification when uploading GGUF models via URL, fixing 500 Internal Server Error crashes. #20263
  • 🚦 Users with Brave Search now experience automatic retry with a 1-second delay when hitting rate limits, preventing failures when sequential requests exceed the 1 request per second limit, though this only works reliably when web search concurrency is set to a maximum of 1. #15134, #20255
  • 🗄️ Administrators with Redis Sentinel deployments no longer experience crashes during websocket disconnections due to improper async-generator handling in the YDocManager. #20142, #20145
  • 🔐 Administrators using SCIM group management no longer encounter 500 errors when working with groups that have no members. #20187
  • 🔗 Users now experience more reliable citations from AI models, especially when using smaller or weaker models that may not format citation references perfectly. Commit
  • 🕸️ Administrators can now successfully save WebSearch settings without encountering validation errors for domain filter lists, YouTube language settings, or timeout values. #20422
  • 📦 Administrators installing with the uv package manager now experience successful installation after deprecated dependencies that were causing conflicts were removed. #20177, #20192
  • ⏱️ Administrators using custom “AIOHTTP_CLIENT_TIMEOUT” settings now see the configured timeout correctly applied to embedding generation, OAuth discovery, webhook calls, and tool/function loading instead of falling back to the default 300-second timeout. Commit

Changed

  • ⚠️ This release includes a major overhaul of database connection handling in the backend that requires all instances in multi-worker, multi-server, or load-balanced deployments to be updated simultaneously; running mixed versions will cause failures due to incompatible database connection management between old and new instances.
  • 📝 Administrators who previously used the standalone “Markdown (Header)” text splitter must now switch to “character” or “token” mode with the new “ENABLE_MARKDOWN_HEADER_TEXT_SPLITTER” toggle enabled, as document chunking now applies markdown header splitting as a preprocessing step before character or token splitting. Commit, Commit, Commit
  • 🖼️ Users no longer see the “Generate Image” action button in chat message interfaces; custom function should be used. Commit
  • 🔗 Administrators will find the Admin Evaluations page at the new URL “/admin/evaluations/feedback” instead of “/admin/evaluations/feedbacks” to use the correct uncountable form of the word. #20296
  • 🔐 Scripts or integrations that directly called Image Generation, Web Search, or Audio APIs while those features were disabled in the Admin UI will now receive 403 Forbidden errors, as backend permission enforcement has been added to match frontend restrictions. #20471
  • 👥 The default group sharing permission changed from “Members” to “Anyone”, meaning users can now share content to any group configured with “Anyone” permission regardless of their membership in that group. Commit

[!TIP]
Looking for an Enterprise Plan? – Speak with Our Sales Team Today!

Get enhanced capabilities, including custom theming and branding, Service Level Agreement (SLA) support, and more!

👏 Massive thanks to our incredible contributors for their hard work and dedication to making this release possible: @ShirasawaSama, @BLACKTHOMAS, @imsamurai, @jannikstdl, @silentoplayz, @joaoback, @rgaricano, @silentoplayz, @Classic298, @majiayu000, @Ithanil, @vincentsunx, @landerrosette, @BarerRocket2678, @rgaricano

Sponsors 🙌

🚀 We’d like to extend a heartfelt thank you to our amazing sponsors for their generous support (Note: We’ve excluded private sponsors from this list. If you’d like to get featured here, feel free to reach out to us!)

@amd, @roosi-gmbh, @rndmcnlly, @feddersen-group, @SRKConsulting, @mushmax, @icanbwell, @loitragg, @haka4githu6, @silenceroom, @crizCraig, @GenieDev101, @CubicleGangster, @TOWaB, @Classic298, @comet-ml, @FominVO, @catalyst-net-nz, @JrdnHans, @hkosm, @VeroFess, @joaoback, @walsermedia, @LuxTronic

下载链接