Share
Running a coding assistant entirely on your own machine used to mean giving up most of what made hosted models useful: multi-file context, agent-style edits, and a decent chat interface. That gap has closed a lot. With Ollama and an editor that treats local models as a real provider instead of an afterthought, you can get the same Ask, Plan, and Agent workflow you would get from a hosted model, just running on hardware you control.
Install Ollama from its official site, then pull a model suited to coding. Reasonable starting points as of this year are Qwen2.5-Coder, DeepSeek-Coder, and Llama 3.3 in its smaller parameter sizes. Larger is not automatically better for editor tasks. A well-tuned 7B or 14B coding model often responds faster and is easier to run continuously than a general-purpose 70B model.
Pull a model with a single command in your terminal, for example ollama pull qwen2.5-coder. Ollama handles the download and keeps the model ready to serve locally on port 11434 by default.
In Orbit Editor, open Settings, go to Providers, and add Ollama as a connection. Point it at your local Ollama server address (localhost:11434 by default if you have not changed it) and select the model you pulled. From here, Ask, Plan, and Agent modes work exactly the same way they do with a hosted provider. The difference is invisible in the interface and very visible in your network traffic, because there is none leaving your machine.
Local models are usually slower per token than a hosted frontier model, so mode choice matters more here than with a hosted provider.
Good for local models almost all the time. Short, focused questions about a function or an error message do not need much context and respond quickly even on modest hardware.
Works well for scoping a change before you commit to it. Local models are often good enough at outlining steps even when they are not fast enough to execute a large multi-file agent run comfortably.
This is the mode most sensitive to hardware. Agent runs can involve many tool calls in sequence, and each one waits on model output. On a 16GB machine running a 7B model, agent mode is usable for small, contained tasks. On a workstation with a strong GPU, it is usable for much more.
Try a smaller quantization first. A 4-bit quantized version of a coding model often loses very little accuracy for editor tasks while running noticeably faster. If that is still too slow, drop down a parameter size rather than switching away from local models entirely.
This is expected for anything requiring broad general knowledge or long chains of reasoning. Local models close the gap fastest on narrow, well-scoped coding tasks: refactors, test generation, and explaining unfamiliar code. Save the hardest architectural questions for a hosted provider if you have one connected as a second option.
Many local models default to a smaller context window than their hosted counterparts. Check your model's configuration and, if your hardware allows it, increase the context length setting when you pull or run the model. This has a direct RAM cost, so it is a real tradeoff, not a free setting.
Three reasons come up most often. First, cost: once a model is downloaded, inference is free, which matters if you are iterating constantly. Second, privacy: nothing about your code or prompts leaves your machine, which matters for regulated codebases or client work under strict contracts. Third, availability: no rate limits, no outages, no waiting on a provider's API status page.
None of this means you should run local models exclusively. Most developers who do this well run a hybrid setup: local models for everyday small tasks, a hosted provider for the harder problems. An editor that treats both as interchangeable providers, rather than treating local models as a separate, lesser feature, is what makes that hybrid workflow practical day to day.
If you want to try this setup, download Orbit Editor and connect Ollama from Settings, or read the documentation for the full provider setup guide.
Z.ai’s frontier coding model — long context, strong agents — is available through Orbit Provider at half price for the launch window. Here’s how to use it in Orbit 0.4.0.
Orbit Team
Jul 20, 2026 · 4 min read
Sign in with GitHub, pick a model, and start coding. Orbit 0.4.0 makes Orbit Provider first-class — with live model lists, wallet balance in the editor, and a cleaner Providers tab.
Orbit Team
Jul 20, 2026 · 4 min read
A feature-by-feature look at how Orbit Editor’s provider-agnostic approach compares to Cursor’s bundled model subscription, and which one fits your workflow.
Orbit Team
Jul 12, 2026 · 6 min read
