Skip to content

Your Own Model

Jump to bottom
Adithya Chinnakkonda edited this page Oct 14, 2026 · 2 revisions

Your Own Model

Kural can run a model on your own computer through Ollama: private, free, no account, no internet. Kural runs it with its own engine (not Claude Code). It reads and edits files, searches your project, asks before running commands, and you keep or undo each change, like with Claude.

Set it up

  • Get started → Your own model: Ollama, a model, a test. See Getting Started.
  • Kural Settings → Your own model → Find & download models… (also under More → Find & download models, or the Command Palette: Kural: Find and Download Models) searches Ollama's library. It lists only models that can chat and use tools (the chat needs tools) and that run on your computer (no cloud-only models), shows how much memory each size needs compared to yours, and downloads with a progress bar. You can also use or delete models you have.

Then pick the model under On this computer in the model menu. That section shows once you have a model that can chat. You need Ollama 0.8 or newer.

Which model?

Bigger is smarter but slower and needs more memory. On a laptop, start with something like qwen3:8b (or qwen3:4b with 8 GB of memory). qwen3-coder:30b or gpt-oss:20b need about 20 GB.

What uses it

The model you pick in the chat is used for the chat, Ask, Ctrl+K, Apply, commit messages and plain-words commands in the terminal. Tab Completion has its own, smaller model (a fill-in-the-middle code model) because it needs to be fast.

What needs Claude

  • Multiple Agents
  • Claude Code's connectors, MCP servers, plugins and skills
  • Linking Jira tickets

Settings

  • kural.localModels.contextLength (default 32768): how much text the model can look at once (your conversation, the files it read, Kural's instructions). More needs more memory: a 9B model at 32768 takes about 7.8 GB, and on a 16 GB computer Ollama then unloads Tab Completion's model to make room (measured: docs/benchmarks/memory-2026-10-08.md). 16384 roughly halves the part the context takes.

Tab Completion while this chat answers: Ollama runs both models on the same graphics chip, so Tab's suggestions slow down while the chat answers; if Claude is set up, Claude helps Tab meanwhile (see Tab Completion).

  • kural.tabCompletion.ollamaUrl (default http://127.0.0.1:11434): where Ollama runs (also for the chat).

Home

Ask for a feature

Clone this wiki locally