ONLYOFFICE AI with a local LLM: Docs 10 changes, Ollama or vLLM setup and the right GPU

ONLYOFFICE AI with Ollama on your own GPU

ONLYOFFICE Docs 10 can run its AI tools against a language model on your own hardware, so drafts, contracts and spreadsheets are summarised, translated and rewritten without leaving the company. ONLYOFFICE's release post of 10 October 2026 states that on server builds "you can connect any AI model, including local models for privacy-sensitive use cases", and its help centre connects Ollama and LM Studio by entering their address. For a small team, one workstation with an NVIDIA RTX PRO 4000 Blackwell (24 GB) running gpt-oss-20b is the entry point; a 96 GB RTX PRO 6000 Blackwell Workstation Edition runs the larger gpt-oss-120b for a department.

This guide is for companies that already use ONLYOFFICE, or are choosing it, and want the editor assistant without a cloud model: firms and public bodies that look for a local alternative to Copilot in Word, and IT teams that have to set it up. It covers what changed in version 10, how the editor talks to a local model, what stays in the company, and which graphics card fits which team.

Two terms first. A language model (LLM) is the AI that reads and writes text; "local" means it runs on a computer in your network instead of a provider's cloud. Ollama is an open-source program (MIT licence) that downloads such models and serves them to other applications over the network.

What changed in ONLYOFFICE Docs 10

Version 10.0 is the release where the AI moved out of the open-source package. The release post says: "In response to user feedback, we've removed built-in AI from the open-source versions." It continues: "AI tools are installed separately." The same post lists the tools per editor:

• Documents: summarise long documents, rewrite or simplify paragraphs, translate text on the page, proofread, build a table of contents. Quick AI actions and AI assistants "are currently available only in the document editor".
• Spreadsheets: generate formulas and explain how they work, analyse data, build forecasts and pivot tables, clean up inconsistent data, fill in missing values.
• Presentations: create outlines and slide titles, write bullet points, generate speaker notes and visual ideas.
• PDF: ask questions about the document, summarise it in one click, draft e-mails from its content, redact sensitive information.

The desktop release post of 13 October 2026 adds that the Desktop Editors have no AI preinstalled ("There is no pre-installed AI component"); the AI module is a separate download. It names "local LLMs (running via Ollama, LM Studio or GPT4ALL)" next to the cloud providers.

Administration differs by edition. ONLYOFFICE's post for developers of 21 October 2026 says that "In commercial editions of ONLYOFFICE Docs 10.0, administrators can configure AI providers once in the Admin Panel", and that after this, users "can no longer add their own keys or change which model is used". Two switches remove AI where it is not wanted: editorConfig.customization.ai with the disabled parameter hides it in one editor configuration, and the server setting aiSettings.enable turns it off and removes the AI tools from the plugin list.

One inconsistency is worth knowing. ONLYOFFICE's plugin documentation, read on 30 October 2026, still lists the AI plugin as included by default in the Community Edition and enabled in the Plugin Manager under Background Plugins, and the AI quick-start guide in its help centre says the AI tools are built into the editors with no plugin to install. The 10.0 release post is the newer statement. On a fresh 10.0 Community installation, plan to add the AI tools yourself; the plugin's source is in ONLYOFFICE's public plugin repository under the AGPL-3.0 licence.

Connecting ONLYOFFICE to Ollama, LM Studio or vLLM

ONLYOFFICE's help centre guide for Ollama takes these steps (the page itself names no Docs version):

1. Install Ollama and set the allowed origins before starting it: OLLAMA_ORIGINS=http://*,https://*,onlyoffice://*.
2. Download a model, for example with ollama run.
3. In the editor, open the AI tab, then Settings, then Edit AI models, and click Add.
4. Choose Ollama as the provider, enter http://localhost:11434 and select the model.
5. Assign the model to tasks. The quick-start guide lists AI Chat, Code, Summarization, Translation, Text Analysis, Image generation, OCR and Vision, each with its own model if you want.

The quick-start guide adds that no model is connected by default "for security reasons", and that for local providers such as LM Studio or Ollama you "enter the Base URL instead" of an API key.

From one PC to a team. localhost in the guide means Ollama runs on the user's own computer. For a team, the model runs on one GPU workstation or server, and the Base URL becomes that machine's address. Ollama's FAQ states that Ollama "binds 127.0.0.1 port 11434 by default", so OLLAMA_HOST has to be set before other machines can reach it, and that OLLAMA_ORIGINS controls which browser origins may send cross-origin requests. By our reading of the two guides, the editor in each user's browser or desktop app sends the AI requests itself, so every user's computer must reach the model address. If your ONLYOFFICE Docs is served over HTTPS, publish the model behind HTTPS as well, through a reverse proxy, because browsers restrict plain HTTP requests from HTTPS pages. Ollama's FAQ describes no user authentication, so limit access to the model port with the firewall.

Context length. Ollama sets the default context, the amount of text the model reads at once, by GPU memory: 4k tokens below 24 GiB, 32k from 24 to 48 GiB and 256k from 48 GiB, per its documentation page on context length. By our reading, the driver reports slightly less than 24 GiB for a 24 GB card such as the RTX PRO 4000, so Ollama starts it at 4k. A long contract does not fit into 4k tokens. Set OLLAMA_CONTEXT_LENGTH to what your documents need; Ollama notes that a larger context needs more memory.

Several users at once. Ollama processes one request per model at a time by default (OLLAMA_NUM_PARALLEL is 1) and queues the rest. For a department, a serving engine such as vLLM handles many requests on one card through an OpenAI-compatible interface. ONLYOFFICE connects providers that are not in its list through a small JavaScript file with a name, a base URL, an optional key and an addon such as "v1"; it publishes no vLLM example, so this connection is something to set up and test. Our comparison of vLLM, SGLang, TensorRT-LLM, llama.cpp and Ollama and the step from Ollama on a PC to a company AI server cover the engines.

You can order the workstation alone, the setup of ONLYOFFICE Docs Community Edition, Ollama and a local model alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Which model for which editor task

Editor tasks are short: a paragraph to rewrite, a page to summarise, a formula to explain. Start with a mid-sized model, test it on your own documents, and move to a larger one where long documents or demanding German or legal text need it.

• gpt-oss-20b, Apache 2.0, is a 14 GB download in Ollama with a 128K context window. It is the entry model for a team on a 24 GB card.
• gpt-oss-120b, also Apache 2.0, is a 65 GB download and needs a 96 GB card for a team, as our gpt-oss-120b guide explains.
• EuroLLM-9B-Instruct, Apache 2.0, covers the official EU languages by its card and was tuned with a focus on "general instruction-following and machine translation". Because ONLYOFFICE assigns models per task, it can take the Translation task while another model writes and summarises.

Our overview of open LLMs for European languages compares further models for German, French or Polish text. For whole files with layout, our guide to translating documents locally with the layout kept is the better route than the editor.

Check what the AI writes before it goes out. ONLYOFFICE's desktop post itself describes reviewing the result before finalising it. For spreadsheets, test a generated formula on a few rows whose result you know; a formula can look plausible and still reference the wrong range.

What stays in the company

With only a local model connected, the selected text, the document passages and the answers go from the editor to your GPU machine and back, and no document text goes to a cloud AI provider. Three settings can change that, and the IT team decides them before rollout:

• Cloud providers. The provider list includes, for example, OpenAI, Anthropic, Google Gemini and OpenRouter. ONLYOFFICE describes locking users to approved providers for its commercial editions; in the Community Edition, agree a written rule and block the providers' addresses at the firewall if needed.
• Web search. The quick-start guide lists web search, for example through Exa, as an optional setting; leave it off if queries must not leave the network.
• MCP servers. Commercial editions can connect MCP servers for access to external tools; each one is a further data path to review.

Our guide on local AI and the GDPR explains what own hosting changes for personal data, and your data-protection officer confirms the processing. Nextcloud users find the comparable assistant in our guide to Nextcloud Assistant with a local LLM.

Which GPU for ONLYOFFICE AI, and for how many users

We size with our users-per-card rule from the guide to how many users one RTX PRO 6000 serves: 90 per cent of the memory the driver reports, less 3 GiB for the runtime, minus the model weights, leaves the memory for conversations. The counts are conversations of 32,768 tokens in flight at the same moment with a 16-bit cache, served by an engine that batches requests, such as vLLM. They are not headcount, because colleagues do not all press the button in the same second.

Our estimates by the users-per-card rule, not measurements; weights and cache per conversation (0.75 GiB for gpt-oss-20b, 1.125 GiB for gpt-oss-120b at 32K) from our gpt-oss-120b guide.

Entry. The RTX PRO 4000 Blackwell is a single-slot card with 24 GB at 672 GB/s and 145 W. By our rule it holds 18.5 GiB, gpt-oss-20b leaves 5.7 GiB, and 5.7 ÷ 0.75 gives about 7 conversations, as in our article on 30B MoE models on 24 GB GPUs. That suits a team of the size ONLYOFFICE recommends for its free Community Edition, which its repository describes with "up to 20 recommended" users. For more simultaneous connections, ONLYOFFICE's download page points to its commercial Enterprise Edition, which ONLYOFFICE licenses separately.

Step up. The RTX PRO 5000 Blackwell with 48 GB at 1,344 GB/s holds 40.0 GiB by our rule and leaves room for about 36 conversations with gpt-oss-20b, or space for a second model such as EuroLLM for translation. A 32 GB RTX PRO 4500 lands at about 17.

Mid. One RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB runs gpt-oss-120b with about 19 conversations; on the 72 GB RTX PRO 5000 the model loads with almost nothing to spare, as our RTX PRO 5000 48 GB or 72 GB comparison shows. The Workstation Edition draws 600 W, the Max-Q 300 W. Two Workstation Edition cards are the realistic maximum for that edition; three or four Max-Q belong only in a chassis built for them with strong directed airflow, because Puget Systems (July 2025) found that directly adjacent Max-Q cards have little room to draw air and Exxact (updated March 2026) reports that four adjacent cards throttled without its custom airflow.

Large. A rack server with two RTX PRO 6000 Server Edition cards runs two copies of gpt-oss-120b and serves the editors next to a company chat such as AnythingLLM, AI search across SharePoint and Confluence with Onyx, Paperless-ngx for incoming mail and other departments. By our estimate it draws about 1.8 kW at the wall, as our page on the server with two RTX PRO 6000 Server Edition explains.

The speed of one answer depends on how fast the card reads the weights: 672 GB/s on the RTX PRO 4000, 1,344 GB/s on the RTX PRO 5000 and 1,792 GB/s on the RTX PRO 6000 Workstation Edition. We found no published measurement of ONLYOFFICE's AI tools on these cards; the speed depends on the model, the engine, the length of the text and how many people ask at once.

We deliver workstations with one RTX PRO 4000, RTX PRO 5000 or RTX PRO 6000 Workstation Edition and servers with RTX PRO 6000 Server Edition for ONLYOFFICE AI; tell us how many colleagues use the editors and in which languages, and we propose the tier and the model.

FAQ