ONLYOFFICE Docs 10.0 Released: AI Integrated Across All Editors and Modular Apps

ONLYOFFICE Docs 10.0 Officially Released: AI Fully Integrated and Modular Architecture

ONLYOFFICE Docs 10.0 has been officially released, bringing approximately 50 new features, over 1,000 bug fixes, and patches for around 35 security vulnerabilities. For teams that self-host, embed, or integrate its document editors into their own tech stack, this release sends a clear signal: AI is no longer an add-on option for a single editor, but a foundational capability spanning the entire suite.

The core highlights of this update can be summarized in three main threads: AI tools covering all editors, the suite shifting to a modular application architecture, and a performance overhaul of the spreadsheet engine. Let's break down each one.

AI Tools Across All Editors: From PDF to Presentations

AI capabilities now exist as a dedicated AI Tools tab in every editor. Its key design feature is reading the context of the currently open file — meaning its suggestions are based on the actual document content rather than generic templated output.

Each editor's AI capabilities have different focuses:

• PDF Editor: Ask questions directly about document content and get answers, generate one-click summaries, draft emails based on file content, and automatically redact sensitive information.
• Document Editor: Summarize, rewrite, simplify long paragraphs, translate text in place, proofread and highlight key points, and automatically generate tables of contents, suggest headings, and apply formatting.
• Spreadsheet Editor: Generate formulas and explain how they work, analyze datasets, spot trends, build forecasts and pivot tables, and clean inconsistent data, flag outliers, and fill missing values.
• Presentation Editor: Create outlines and slide titles, write bullet points and build narrative flow, and generate speaker notes and visual ideas.

Actions, Agents, and Saved Prompts

Beyond default prompts, users can extend how they use AI through three building blocks:

Feature	Function
Quick AI Actions	One-off tasks, such as adjusting paragraph tone
AI Agents	Continuously running, responding in real time as you edit
Saved Prompts	One-click reuse of common requests in AI Chat

Note that Quick AI Actions and AI Agents are currently only available in the Document Editor.

Bring Your Own Model: A Key Capability for Self-Hosted Users

For self-hosted users, Bring Your Own Model may be the most significant update:

• Cloud version: Enable ONLYOFFICE AI services by topping up your Wallet.
• Server version: Connect to any AI model, including local models. This is hugely important for privacy-sensitive scenarios where data cannot leave your own infrastructure.
• Both versions: Set a default model and assign different models to specific tasks. For example, use one model for summarization and another for generating formulas.

One change worth noting: based on community feedback, built-in AI has been removed from the open-source version. AI tools are now installed separately, keeping the base installation package lean.

---

The technical foundation behind Bring Your Own Model is a model-agnostic API abstraction layer. ONLYOFFICE interfaces with different inference backends through OpenAI-compatible interface specifications (such as /v1/chat/completions), so as long as a model service exposes a standard interface, it can be connected whether it's a cloud-based closed-source model or a locally deployed open-source model. For local deployment, common options include Ollama, vLLM, LocalAI, or LM Studio, which can expose weights from Llama, Mistral, Qwen, etc. as inference services on a local network.

Three things to watch in self-hosted scenarios: first, context length — when processing long document summaries, the model's maximum token limit directly determines whether it can ingest the full text in one pass; second, inference hardware — 7B-class models typically require 8-16GB of VRAM, while 70B-class models need multiple GPUs or quantization; third, network isolation — if the model and ONLYOFFICE services are not on the same trusted network, internal addresses or mutual TLS authentication should still be used to keep data under control. This also explains why removing built-in AI from the open-source version makes sense — it returns model choice and deployment costs to the operations team rather than forcing a cloud service subscription.

Modular Apps: The New Docs + Apps Architecture

Version 10.0 replaces the previously fixed single workspace with freely composable modular apps:

• AI agents
• Forms
• Rooms
• Files

This design's logic is straightforward: let software shape itself around workflows, not the other way around. Enable only the apps you need and skip what you don't. For integrators, the configuration process is three steps: install the editors, extend apps as needed, and integrate the package into your own solution.

---

The engineering implication of a modular application architecture is on-demand assembly of frontend and backend capabilities. A traditional monolithic suite loads all routes and resources at startup, whereas with modularization each app (AI agents, Forms, Rooms, Files) has its own entry point and dependency declarations, so installation package size and startup memory scale with what's enabled.

For integrators, this architecture is typically controlled through manifest files or feature flags: choose which modules to register at deployment time, then deliver the corresponding interface at runtime based on user permissions. The benefit is that upgrades can be rolled out gradually per app, and a problem in one module won't drag down the entire workspace; the cost is that cross-module state sharing requires additional interface contracts — for example, the permission model between Files and Rooms must be explicitly bridged. It's therefore better suited to teams that already have a clear microservices or plugin system, rather than those looking to treat the suite as a black box for wholesale replacement.

PDF Editor: Stronger Control and Privacy Protection

For users handling sensitive files, the PDF Editor gets a focused set of upgrades:

• Remove metadata with Redact: Strip hidden information before sharing files (Redact tab → Remove hidden information).
• Automatic table recognition: Automatically detect tables when using Edit Text on a PDF, with no need to manually rebuild rows and columns (Edit PDF).
• Reusable signatures in forms: Save a signature once and call it from a list anytime (Forms → Signature fields).
• Sign directly from the comment tab: Complete signatures right where you're working.
• Customizable annotations: Adjust borders, backgrounds, and transparency.
• Form spell-check toggle: Enable or disable per form field (Forms → Field settings → Check spelling).

---

Metadata is an easily overlooked privacy leak surface in PDF files, with typical fields including author, creation software, modification time, GPS coordinates, and even residual objects from deleted text. The Remove hidden information feature under the Redact tab traverses the document object tree to clear these invisible properties and content that has been marked as deleted but still exists in the data stream.

Note that simply covering text with a black rectangle is not safe, because the underlying characters can still be copied or extracted — true content deletion must be performed rather than visual masking. Additionally, automatic table recognition relies on implicit text positions and line vectors in the PDF, and is often ineffective for scanned documents or image-only PDFs, which still require OCR preprocessing. Reusable signatures are stored in encrypted form locally or in form-level configuration and are not exported when shared with others, reducing the risk of signature misuse.

Document Editor: Collaborative Editing and RTL Support Improvements

Follow collaborator cursors is a highlight feature of this Document Editor update. During real-time collaboration, you can jump directly to a teammate's cursor position and follow their editing process, making pair review of spec documents or README drafts much smoother (Users who are editing the file → Follow / Stop following).

Right-to-left (RTL) support has also improved. Graphic formulas and other inline objects behave more reliably in RTL documents, which is good news for teams publishing documents in Arabic, Hebrew, and other RTL languages.

Other formatting improvements include:

• Keep with next now supports tables, as well as paragraphs and tables within block content controls
• Cross-page line breaking lets long tables flow naturally
• Horizontal line settings window provides fine-grained control
• Line-based clipping handles content exceeding line height

Spreadsheet Editor: Under-the-Hood Performance Overhaul

Performance is the core theme of this version's spreadsheet, with two key engineering changes worth noting:

First, separation of style and data storage. Formatting and cell values are stored independently. Second, limiting the cell traversal range. Each operation only processes the cells that are actually needed.

The practical result is reduced wait times and smoother handling of large workbooks. Conditional aggregation functions that teams commonly use have also been optimized:

SUMIF   AVERAGEIF   COUNTIFS   SUMIFS
AVERAGEIFS   MAXIFS   MINIFS

As a result, heavy formulas like the following should complete faster on large tables:

=SUMIFS(D2:D50000, A2:A50000, "EMEA", C2:C50000, ">=2026-01-01")

There are also multiple usability upgrades: function constant hints showing constant positions as you type; formula range highlighting for quick reference verification; animated selection boxes showing exactly what will be copied or cut; support for resizing all cells at once with the mouse; an autocomplete toggle for cell values (File → Advanced Settings); and worksheet backgrounds preserved across open and save cycles.

---

Separation of style and data storage is a classic storage model optimization. Traditional spreadsheets often mix formatting information (fonts, borders, fills) with cell values in the same structure, causing every recalculation or redraw to process large numbers of value-irrelevant properties. After 10.0 splits the two into independent storage, data changes only touch the value layer while the style layer remains lazily read, reducing memory copies and serialization overhead.

Limiting the cell traversal range corresponds to the sparse matrix concept: the actual effective area of a spreadsheet is often far smaller than the theoretical maximum row count (~1.04 million rows in older versions), and old implementations might scan the entire range. The new version only traverses blocks that have content or formula dependencies, reducing complexity from O(rows×columns) to O(actually used cells). This benefits aggregation functions like SUMIFS that need to scan entire columns most noticeably — the formula semantics are unchanged, but scan boundaries are dynamically narrowed by the engine based on data density, so the conditional sum over a 50,000-row range in the example can be significantly accelerated.

Presentation Editor: New Views and GIF/Video Export

The View tab adds several new display modes: Outline View, Reading View, Notes Page, Notes Master, and Handout Master.

Export to GIF or video lets presentations be turned into shareable short clips, suitable for announcements, demos, or social posts (File → Download as). Print settings have also been expanded to print notes, handouts, and outlines, and can include hidden slides, backgrounds, borders, comments, and slide numbers, scaling content by page.

Format, Plugin, and Security Updates

On the format support side, opening and converting .ofd files has been added, along with saving spreadsheets in the classic .xls format.

Other new features include: SVG-to-editable-shape conversion, chart data table support, and a redesigned plugin manager (with a left-side category panel). New community plugins cover spelling and grammar checking (Antidote), reusable templates (Curly), Chinese conversion and formatting tools, and data import/export between Salesforce and spreadsheets.

On the security front, this version fixes around 35 vulnerabilities. The original text notes that AI is giving attackers new ways to discover and exploit weaknesses, so security was placed at the core, making the suite more resilient when teams handle confidential data at scale.

What It Means for Developers and Integrators

The direction of ONLYOFFICE Docs 10.0 reflects a trend: document suites are shifting from feature stacking to architectural flexibility and AI-native experiences. The Bring Your Own Model capability is particularly worth evaluating for self-hosted teams — it returns control of AI integration to the deployer rather than binding it to a single cloud service. Modular apps reduce integration complexity, letting developers assemble what they need.

Are you planning to connect a local model? Which feature is most worth trying? Share your thoughts in the comments.