AI in Nextcloud with a local model: the options in 2026
· By Bee Flow · 4 min read
Three ways to give Nextcloud a model on your own hardware: llm2, Nextcloud's OpenAI integration pointed at Ollama, or Bee Flow with a local runtime. What each needs, and what to get right during set-up.

In Nextcloud you choose the model behind the AI features. The Assistant schedules a task through the Task Processing API, and whichever provider app is registered for that task type does the work. If that model has to run on your own hardware, an administrator has three practical routes in 2026. The Nextcloud facts below come from the Nextcloud 35 manuals, the current release as we write this.
llm2: the model inside Nextcloud
llm2 is an ExApp, so Nextcloud runs it as a container through AppAPI (version 3.1.0 or later). It uses llama.cpp and loads GGUF files. It ships with Qwen 3.5 9B, Gemma 4 E4B and OLMo 3 7B Instruct, and you can add other GGUF models by copying them into the container's data directory. The documented hardware is an NVIDIA GPU with at least 8 GB of VRAM and CUDA 12.4 or later, 12 GB of system RAM and a CPU with AVX and AVX2. For heavier workloads, the manual points to an OpenAI-compatible service.
In Nextcloud's Ethical AI Rating, llm2 is the green option for text processing.
integration_openai pointed at your own runtime
The OpenAI and LocalAI integration talks to any service with an API similar to OpenAI's. Its list of supported providers gives http://localhost:11434/v1 as the default address for Ollama. You enter the base URL without the endpoint path, and the app reads the model list from the service's /models endpoint. The runtime can live on any machine with a GPU, apart from Nextcloud.
The AI overview rates integration_openai with Ollama yellow. The AI as a Service page describes how Nextcloud tests the Assistant tasks of this app, with OpenAI models against the OpenAI API. With your own runtime, try the tasks you plan to use on the model you chose.
Bee Flow with a local runtime
In our setup the model call comes from the Bee Flow server, a separate stack next to Nextcloud. The Bee Flow app for Nextcloud is a small ExApp that shows Bee Flow inside Nextcloud and passes requests on. Bee Flow has one adapter for self-hosted runtimes and knows Ollama, vLLM, llama.cpp, LM Studio, SGLang, LocalAI, Text Generation Inference, Jan and KoboldCpp by name, plus any other OpenAI-compatible endpoint. Under Admin → AI Config → API Keys, Find running runtimes probes the default ports, and variables such as OLLAMA_URL or VLLM_URL register a runtime at start-up. The local models docs have the details, and our Ollama post walks through the whole stack.
The connector also registers Bee Flow as a Task Processing provider for the text task types. Nextcloud's own AI features can then run on that same local model. Image generation, speech and transcription are not registered.
| llm2 | integration_openai + runtime | Bee Flow + runtime | |
|---|---|---|---|
| Runs as | ExApp container | Nextcloud app that calls the runtime over HTTP | Separate Bee Flow stack, plus a connector ExApp |
| Model runtime | llama.cpp inside the container | Ollama, LocalAI or another OpenAI-compatible server | Nine named runtimes or any OpenAI-compatible server |
| Hardware in the docs | NVIDIA GPU with 8 GB VRAM, 12 GB RAM | Whatever the runtime needs | 2 cores and 4 GB for Bee Flow (about 7 GB more with the PII guard and classifier), plus the runtime |
| Where users work with it | Nextcloud Assistant | Nextcloud Assistant | The Bee Flow app in Nextcloud, and the Assistant through the provider |
Getting the set-up right
Start with workers. Nextcloud picks up AI tasks through background jobs, which run every five minutes by default. To have tasks picked up straight away, the AI overview recommends four or more dedicated workers. Since Nextcloud 32.0.7 the command is:
sudo -E -u www-data php occ taskprocessing:worker -v -t 60
You run it in a loop, or as the nextcloud-ai-worker@ systemd unit the manual describes. Older versions use background-job:worker. If the Assistant is installed and still answers slowly, Nextcloud Assistant slow or not answering goes through the causes in order.
Then there is localhost. Ollama binds to 127.0.0.1:11434 by default, according to its FAQ, and localhost inside a container is that container. When Nextcloud runs in Docker, as All-in-One does, the default URL points at the Nextcloud container itself. Use the host's address and start Ollama with OLLAMA_HOST=0.0.0.0, behind a firewall, since these runtimes have no authentication of their own. The same goes for Bee Flow in Docker, where our docs suggest host.docker.internal.
The model selected in Nextcloud's settings has to be one the runtime serves, so pull it there first. Ollama names a missing model in its error, for example model 'gpt-3.5-turbo' not found, try pulling it first.
Agent features need tool calling. The Context Agent documentation says the model must support it, and the llm2 page says which of its bundled models support tools. Some runtimes need a flag as well. For Bee Flow, llama.cpp needs --jinja, and vLLM needs --enable-auto-tool-choice with a tool-call parser. Without them the model answers in prose where it should call a tool.
Taking the Bee Flow route
With Bee Flow, the model on your own hardware can take the Assistant's text tasks, and the same runtime serves chat, agents and questions to knowledge bases you fill from your Nextcloud folders. Each model tier can get its own model, so a small local one can take short jobs while a larger one does the rest. Installing Bee Flow next to Nextcloud covers the connector and the reverse-proxy rule that HaRP needs, AI for Nextcloud shows what your users get, and self-host or EU cloud covers running Bee Flow itself. If you want to talk through which route fits your hardware, contact us.