Local coding with Continue and Ollama
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
A small personal site, a Python script that sorts photos, or a Home Assistant add-on is a different job from shipping software for a company. At home the stakes are usually two: save time on repetitive edits, and keep bits of code off a cloud assistant. Those bits sometimes include a password in a comment, a test API key, or personal data.
In short: goal, who it is for, what you need first
- Goal: point the Continue extension at Ollama so you get a code chat and autocomplete inside VS Code, with the project staying off a hosted coding service.
- Who: a hobby developer, a student, or a parent editing personal scripts. A company monorepo is a different job.
- Before you start: Ollama installed (Windows, Linux, and Mac or the Windows walkthrough), a 7B–13B code model downloaded, 16–32 GB of RAM or a GPU with enough VRAM (VRAM and model size), and a current VS Code.
Words you will see in the UI: Continue and VS Code say inline, autocomplete, and diff. Here that means suggestions while you type, automatic completions, and the before/after comparison of an edit.
Continue is an open-source extension for Visual Studio Code (and JetBrains). It adds a chat that can see the file you are in, edits from a written instruction, and suggestions while you type (inline in the interface). Ollama is the engine: it loads models on your machine and exposes a local API (an HTTP interface the editor uses to talk to the model), usually on port 11434. Point Continue at that base URL, the apiBase field, for example http://localhost:11434. Prompts and snippets then stay on the home network.
What Continue does, and what Ollama adds
Continue is a conversation layer on top of the editor. Your compiler and Git still do their own jobs. You can select a file and ask “explain this function” or “add basic unit tests,” request a change across several files, or accept a suggestion while you type. The model sees the context you send: the open file, the selection, and sometimes other indexed files, depending on your settings.
Ollama downloads model weights, serves them from a background service, and answers chat requests in an OpenAI-compatible shape. Continue connects with apiBase (the base URL of the Ollama API) and the exact model name from ollama list. While Ollama listens only on the machine, and port 11434 stays closed to the public internet, the traffic stays on your network. For a hobby project, that is a strong privacy argument. A huge company repository is outside the aim of this page. See privacy when the AI stays home.
If Ollama is still missing, start with installing Ollama on Windows, Linux, or Mac or the Windows walkthrough. A first test with VS Code still closed: a first local AI in about 15 minutes.
GGUF, quantization, and 7B–13B code models
The models Ollama fetches are often packaged from GGUF files. GGUF is the format associated with llama.cpp (Georgi Gerganov Unified Format: compressed tensors for CPU and GPU inference). In practice it stores a quantized model (Q4_K_M, Q5_K_M, Q8_0, and so on). Quantization lowers the numeric precision of the weights so they fit in VRAM or system RAM, with a small loss of sharpness. For hobby code, Q4_K_M on a 7B or 8B “coder” model is the usual compromise. A 13B at Q4 needs more memory and follows longer refactors more steadily.
You can skip hand-editing GGUF files. ollama pull downloads a ready-made Modelfile. Prefer instruct variants aimed at code (names that often include coder or code, or families with a reputation for development). The Ollama catalog moves quickly. In October 2026, 7B–13B is still the realistic band for a home development PC with a single graphics card. A 70B from the front page of a catalog will fall back to a slow CPU on a mini PC, and autocomplete will feel stuck. VRAM and which model and choosing a first model line up size, quantization, and how fluid the editor feels.
Example commands (match the tags to the catalog that day):
ollama pull qwen2.5-coder:7b
ollama pull deepseek-coder-v2:16b-lite-instruct-q4_K_M
ollama run qwen2.5-coder:7b
Try them in a terminal. If replies drag, or the process spills entirely into system RAM, drop a size, or move to a machine with more RAM: 32 vs 64 GB for language models.
Procedure: Ollama ready, then Continue in VS Code
Prepare Ollama and check the API
- Install Ollama and let the service start (the taskbar icon on Windows,
systemdon Linux). - Download a code model that fits your RAM and VRAM (see the model section).
- Check the API:
curl http://localhost:11434/api/tagsshould list your models.ollama listis the other check. - Optional: if Ollama runs on another computer in the house, note its reserved DHCP address and allow port 11434 only on the home network. Leave the router’s WAN forward closed.
Install Continue in VS Code
- Open VS Code, the Extensions view, search for Continue (publisher Continue), and install the official extension.
- Open the Continue panel (the sidebar icon), then Settings for the model. Depending on the version that is a
config.yamlfile or a settings screen.
Point the model at Ollama and apiBase
- Add an Ollama model with:
provider:ollamamodel: the exact name fromollama list(for exampleqwen2.5-coder:7b)apiBase:http://localhost:11434, orhttp://IP_OF_THE_SERVER:11434if Ollama is on another machine on the home network
- Set that model as the chat model and, if the screen allows it, as the autocomplete model (suggestions while you type).
- Restart VS Code if the change does not show up.
Continue configuration (docs as of October 2026)
The syntax can change with Continue. Check the official documentation before you copy. The keys stay in English (provider, model, apiBase, roles): Continue requires that spelling. Copy the block line by line, with no spaces before models:
models:
- name: Ollama Code 7B
provider: ollama
model: qwen2.5-coder:7b
apiBase: http://localhost:11434
roles:
- chat
- autocomplete
With Docker (Ollama and Open WebUI in Docker), the host can be http://127.0.0.1:11434 mapped from the container. Test with curl from the same computer that runs VS Code.
Try Continue in the editor
- Open a local project. A GitHub remote is unnecessary.
- Select a function, open the Continue chat, and ask for an explanation in plain English.
- Ask for a focused change (“add error handling when this file opens”) and read the before/after comparison (the diff in Continue) before you accept it.
- Turn on suggestions while you type (inline) if they are available. If they are slow, shrink the context or switch to a lighter 7B model.
- Leave secrets out of the chat. Even on a local machine, logs and screenshots can leak. Get used to stripping them out.
Connection recap
apiBase: http://localhost:11434
Default Ollama port: 11434
model: the exact name copied from "ollama list"
If Ollama runs on another machine on the home network, replace localhost with that server’s fixed address in apiBase only. Keep the firewall tight, and leave port 11434 unforwarded on the router.
Hardware: mini PC, GPU, and VRAM for code
Once Continue is connected, a slow autocomplete on a light 7B is a hardware question. Confirm the three lines in the recap above before you replace the computer. In one sentence: a quantized 7B often fits a 32 GB mini PC with an 8 GB RTX. A comfortable 13B wants about 12–16 GB of VRAM, or a Mac with plenty of unified memory. Longer guides: VRAM and which model, Mac mini M4, mini PC comparison, and the hardware selector (use, noise, and budget, with a specific product left to you).
Household ranges for October 2026. Measure on your machine (ollama ps, the system monitor, or nvidia-smi on an NVIDIA GPU).
| Profile | VRAM / RAM | Model that fits | Comment |
|---|---|---|---|
| 16 GB laptop | Integrated graphics, RAM under pressure | 3B–7B Q4 | Fine for learning. Autocomplete can stutter |
| 32 GB mini PC + RTX 4060 8 GB | 8 GB VRAM | 7B–8B Q4/Q5 | A solid price-to-comfort balance for a desk |
| 64 GB workstation + RTX 4070 12 GB | 12 GB VRAM | 7B feels easy, 13B Q4 | Multi-file refactors hold together better |
| Mac mini, 24–32 GB unified | Unified memory | 7B–8B | Quiet. Check the Ollama tags for Apple |
Continue can run on the developer’s laptop while Ollama runs on a desktop in the living room (a quiet mini PC). Same apiBase rule, with the home-network address. For a machine left on all day: noise and power of a mini PC, electricity budget. Keep models on an SSD (SSD storage for models). External and integrated graphics: GPU, iGPU, eGPU, and OCuLink.
Optional hardware purchase, after a test shows that VRAM or RAM is the bottleneck. Check the current price (October 2026). LocAIhome does not print a fixed price here. The store shows the amount. Search Amazon to compare what is listed now: mini PCs for a local model on Amazon.com, RTX 4060 cards on Amazon.com. Partner links (rel="sponsored", tag roboteducatif-20). The disclosure is on the affiliate page.
For chores you would rather automate than edit in an IDE, the n8n and local model path sits next to this one.
Honest limits, and other tools
A local 7B–13B gets obscure APIs wrong, misses very new libraries, and loses the plot on a huge project that was never indexed. Continue is less deeply wired into the Microsoft ecosystem than Copilot. To compare hardware cost with cloud subscriptions over three years, open Ollama vs ChatGPT over 3 years.
LM Studio (GPU offload guide) is handy for trying GGUF files before you serve them with Ollama. If you only want a conversation, with the editor closed, Open WebUI is enough. The private family assistant often shares the same machine. Household automations go through n8n and a local model or Home Assistant rather than through Continue.
Security and everyday habits
Leave Ollama off the public internet unless you have strong authentication in front of it. Update Ollama and Continue on a regular schedule. On a PC shared with a teenager, set a rule: an employer’s code stays off the family model. Local inference lowers the chance of a leak to a cloud vendor. It does nothing about hallucinations. Read generated code before you run it, especially when it touches the network or the file system.
In short: install Ollama, pull a 7B–13B code model, configure Continue (the recap is apiBase / 11434 / ollama list), try it in the editor, then change the hardware or the model if the delay gets in the way. Hardware overview: choosing hardware for home AI. Software overview: the local AI software hub.
Common questions
Which coding model should I pull in Ollama?
Aim for an instruct model built for code, in the 7B to 13B range, quantized Q4_K_M. That is enough for personal scripts when the VRAM (the graphics card’s memory) or unified memory can hold it. See the LocAIhome VRAM guide.
Does Continue work without Ollama?
Yes, with other model servers. This guide uses local Ollama so prompts and snippets from the project stay on your machine.
How close is this to GitHub Copilot?
The shape is similar: autocomplete plus a chat in the editor, with less built-in product integration and more settings to tune. For a household, the point is that the code stays home.
Sources
Narrow it down with the hardware selector · French version of this section