Continue is an open-source coding assistant that can run in Visual Studio Code and talk to a model on your own PC. Paired with Ollama, that means chat, edits and autocomplete can stay local, which is the point if you do not want source code sent to a cloud API.
Forge PC has not benchmarked Continue, and we have not ranked coding models against each other. The install, configuration and troubleshooting steps below come from Continue's official "Using Ollama with Continue" guide.
You need Ollama installed and running first. Our Ollama and LM Studio install guide covers that step. This page starts from a working Ollama install on port 11434.
Prerequisites
Continue's Ollama guide lists:
Windows, macOS or Linux
8GB of RAM minimum, 16GB or more recommended
At least 10GB of free storage
The Continue extension installed
It also gives rough memory bands for models: about 4GB of RAM for 1.5B–3B models, 8GB for 7B, 16GB for 13B and 32GB for 32B. Those are Continue's figures, not measurements we have taken. If a 32B model will not load, use a smaller one. Our local AI RAM and VRAM guides explain the same constraint in more detail.
Install a model the way Continue expects
Start Ollama, then pull a model. Continue is explicit: use ollama pull, not ollama run, because run opens an interactive session you do not need.
ollama pull qwen2.5-coder:1.5b
ollama list
The tag matters. Continue warns that ollama pull deepseek-r1 installs :latest, which may not be the :32b version a config block expects. Pull the exact name you will put in the config. If Continue later reports 404 model "something" not found, try pulling it first, the names do not match.
Continue's guide mentions several starting points, depending on the job:
Autocomplete: qwen2.5-coder:1.5b or starcoder2:3b
Code generation: qwen2.5-coder:7b, codellama:13b or deepseek-coder:6.7b
Chat and reasoning: llama3.1:8b, mistral:7b or deepseek-r1:32b
You do not have to install all of them. Start with a small coder model if you are short on memory.
You can check that Ollama itself is up with curl http://localhost:11434. The official check is that it should report that Ollama is running.
Connect Continue to Ollama
Continue documents three ways to point the extension at local models.
Autodetect is the least fiddly. In YAML:
models:
- name: Autodetect
provider: ollama
model: AUTODETECT
roles:
- chat
- edit
- apply
- rerank
- autocomplete
You can also choose Autodetect in the model selector in the GUI. Continue will scan ollama list and offer what it finds. You can set apiBase to another machine's address if Ollama is running elsewhere.
Manual configuration is useful when you want roles split, for example a larger model for chat and a small one for autocomplete:
models:
- name: Local chat
provider: ollama
model: llama3.1:8b
apiBase: http://localhost:11434
roles:
- chat
- edit
- name: Local Autocomplete
provider: ollama
model: qwen2.5-coder:1.5b
apiBase: http://localhost:11434
roles:
- autocomplete
The model value must match ollama list exactly.
Model blocks in config.yaml (for example uses: ollama/qwen2.5-coder-7b) only provide configuration. Continue still expects you to have pulled the matching tag locally.
Reload the configuration after you save it, and enable tab autocomplete in VS Code if that is the feature you want.
If something fails
Continue's own troubleshooting list covers the usual cases:
"Model requires more system memory to run." Continue may use a higher default context length than other tools. Reduce contextLength in the config (the guide suggests trying 2048) or use a smaller model.
404 model not found. Pull the exact tag, then check ollama list.
"Agent mode is not supported." Tool use is required for Agent mode. You can add capabilities: [tool_use] in the config, but Continue notes that some models still will not work, including cases it has seen with DeepSeek R1. It suggests models with confirmed tool support, such as Llama 3.1 or Mistral, or turning on system-message tools.
Connection problems. Confirm curl http://localhost:11434 works, and that port 11434 is not blocked. For a remote Ollama host, Continue mentions setting OLLAMA_HOST=0.0.0.0:11434 on that machine.
Use ollama ps to see what is loaded and how much memory it is using, and ollama logs if you need to debug the server.
Keep this on your own PC unless you have deliberately set up remote access. A local coding assistant is only private for as long as the model and the API stay local.
A note on config files and privacy
Continue's guide shows a config.yaml with a schema version and a model list. The exact on-disk path can vary with the Continue version and whether you use a local or shared config, so if a file you edited seems to have no effect, check the path the extension itself reports rather than guessing. Reload the configuration after every save.
The attraction of this setup is that prompts and source code can stay on the machine that runs Ollama. That only holds if the model, the API and the editor are all local. Pointing Continue at a remote apiBase or at a cloud provider is a different design. If privacy is the reason you chose Ollama, leave apiBase at http://localhost:11434 and do not paste snippets into a second, cloud-hosted assistant "just to compare".
Once chat works, try one small edit or autocomplete suggestion in a throwaway file before you rely on it in a real project. Continue can be helpful and still be wrong.
Our verdict
Install Continue, pull one Ollama model whose name you will actually use, and point the extension at http://localhost:11434 with Autodetect or a short config.yaml. Start with a small coder model if memory is tight, and treat Agent mode as optional and model-dependent. We have not tested Continue's output quality; the steps come from its official Ollama guide.
