Skip to main content

Working with AI Models

Libre WebUI can use local Ollama models and plugin-backed cloud models in the same workspace. The Model Manager shows installed Ollama models, running models, live Ollama Library results, Hugging Face GGUF entries, and Ollama Cloud entries where available.

Choosing a First Model

Use these as starting points, then switch based on your hardware and task:

ModelGood forTypical fit
gemma3:4bFast general chat8 GB RAM or entry GPU
qwen3:8bGeneral chat, coding, multilingual use16 GB RAM or 6-8 GB VRAM
deepseek-r1:8bReasoning-heavy prompts16 GB RAM or 6-8 GB VRAM
mistralBalanced general assistant work8-16 GB RAM
nomic-embed-textDocument embeddingsSmall local embedding model

Large models such as 30B, 70B, and MoE models can be excellent, but they need much more memory. If you are not sure, start small and move up after the model is working smoothly.

Model Manager

Open Models from the sidebar to:

  • Pull models from Ollama by name.
  • Search the live Ollama Library instead of relying on a static list.
  • View installed and running models.
  • Stop or unload running models.
  • Delete models you no longer need.
  • Pull Hugging Face GGUF models through Ollama when compatible.
  • Pull Ollama Cloud models from the cloud filter.

For Ollama Cloud results, the UI normalizes cloud model names before pulling. If a cloud model requires the :cloud or -cloud suffix, Libre WebUI applies that for you from the cloud model flow.

Local vs Cloud Models

ModeStrengthsTradeoffs
Local OllamaPrivate, offline after download, predictable costDepends on your CPU/GPU/RAM
Ollama CloudFamiliar Ollama workflow without local hardware limitsRequires cloud access and network
Provider pluginsAccess to managed models from multiple providersAPI keys, provider pricing, and provider privacy policy apply

You can keep local models for private work and enable provider plugins for tasks that need larger hosted models.

Models for Work

Work needs a chat model that can call tools. It can use:

  • An installed Ollama model that advertises the tools capability.
  • An Ollama Cloud model available through the configured Ollama endpoint.
  • A model listed by an active chat or completion plugin with credentials configured for the current administrator.

Plugin-backed Work runs use the provider adapter appropriate to the configured plugin: OpenAI-compatible, Anthropic, or Gemini. Libre WebUI persists the exact provider type and plugin identifier with the task and each run, so a plugin cannot capture an identically named Ollama model. If the selected model or provider rejects tool calling, the run fails instead of silently switching to another provider.

Local Ollama keeps model requests on the configured Ollama infrastructure. With a remote model, the configured provider receives the Work system prompt, conversation, tool definitions, and tool results. Tool results can include source text, command output, or directory listings requested by the model. Workspace volumes and provider credentials remain on the backend host, but a file's contents can leave that host when they are included in a tool result.

One autonomous Work run can make multiple model calls. Check the remote provider's pricing, retention, and training policies before using sensitive projects. Libre WebUI shows a remote-provider notice in Work with a per-user dismiss control.

Hardware Guide

SystemPractical model rangeNotes
CPU only, 8-16 GB RAM1B-4BGood for light chat and testing
8 GB VRAM4B-8B quantizedComfortable starting point
12-16 GB VRAM8B-14B quantizedGood daily driver range
24 GB VRAM14B-32B quantizedStrong local workstation
48 GB+ VRAM or large unified memory32B-70B quantizedLarge model experimentation

Quantized models use less memory. Q4 quantizations are usually the practical default; Q8 uses more memory for better quality.

Task-Based Recommendations

TaskModel direction
Fast chatgemma3:4b, mistral, small Qwen models
CodingQwen Coder, DeepSeek Coder, Codestral, provider coding models
ReasoningDeepSeek-R1 family, larger Qwen models, provider reasoning models
VisionMultimodal models such as LLaVA, Qwen VL, or provider vision models
Document searchnomic-embed-text or another embedding model
Text-to-speechTTS plugins such as Qwen3-TTS or Kyutai TTS

Provider model names change frequently. In Libre WebUI, use the provider’s model discovery where available, or paste the exact model ID from the provider dashboard.

Prompting and Settings

  • Lower temperature (0.1-0.3) for factual, repeatable answers.
  • Medium temperature (0.5-0.7) for normal assistant work.
  • Higher temperature (0.8+) for brainstorming and creative writing.
  • Keep context length reasonable when you are close to memory limits.
  • Use personas when you want persistent model parameters and a reusable system prompt.

Troubleshooting

Pull fails

  • Confirm Ollama is running: ollama list.
  • Try the same pull in a terminal to see Ollama’s raw error.
  • Check disk space before pulling large models.
  • If you are using the Model Manager cloud filter, let Libre WebUI handle cloud suffixes.

Responses are slow

  • Try a smaller model or a lower quantization.
  • Check ollama ps to see what is loaded.
  • Close other GPU-heavy apps.
  • Reduce context length.

Out of memory

  • Move from Q8 to Q4.
  • Use an 8B model instead of a 14B model.
  • Keep only the model you need loaded.
  • On Docker, confirm the container can reach the GPU or the host Ollama instance.