DEV Community

niuniu
niuniu

Posted on

I Replaced GitHub Copilot with Continue.dev + a Local Model for 60 Days — My Code Didn't Get Worse, My Wallet Got Happier

I cancelled my $10/month GitHub Copilot subscription two months ago. Not because of the money — $10 is coffee money — but because I realized I was sending every keystroke in proprietary client code to a cloud API, and the contract I just signed had opinions about that.

So I went full local: Continue.dev (open-source VS Code/JetBrains extension) + Ollama running Qwen2.5-Coder-7B on my RTX 4060 8GB. Here's what 60 days of real usage looked like.

The Setup (15 minutes, $0) 

# 1. Ollama
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:7b

# 2. Continue.dev — install from VS Code marketplace, then ~/.continue/config.json:
{
  "models": [{
    "title": "Qwen Coder Local",
    "provider": "ollama",
    "model": "qwen2.5-coder:7b"
  }],
  "tabAutocompleteModel": {
    "title": "Qwen Autocomplete",
    "provider": "ollama",
    "model": "qwen2.5-coder:1.5b"
  }
}
Enter fullscreen mode Exit fullscreen mode

Trick: use the 7B model for chat/edits but the 1.5B for tab autocomplete — autocomplete needs to fire in <200ms or it feels broken, and the small model does ~45ms on a 4060.

The Numbers I Tracked

I logged every suggestion for 60 days (Continue has a built-in dev data export):

Metric Copilot (before) Continue + Qwen (local)
Tab suggestion acceptance rate 31% 24%
Median autocomplete latency ~180ms (network) ~45ms (local)
Chat "good answer" rate (my judgment) ~85% ~70%
Cost / 60 days $20 $0 (+~$3 electricity)
Keystrokes leaving my machine 100% 0%

Where Local Actually Won

  1. Latency. 45ms local vs ~180ms network means autocomplete appears before I finish thinking, not after. This is underrated — it's the difference between "tool" and "annoyance".
  2. Airplane mode. I wrote a full FastAPI feature on a train with no wifi. Copilot is a brick offline.
  3. Privacy review. Client's security team asked "does your tooling exfiltrate code?" Answer went from a paragraph of caveats to "no."

Where It Lost (Honest Part)

  1. Big refactors. "Rename this concept across 40 files and update the tests" — Copilot's larger models handle the cross-file context noticeably better. Local 7B starts hallucinating file contents after ~4 files.
  2. Obscure APIs. Asked both to write code against a niche payment provider's SDK. Copilot knew it. Qwen confidently invented methods that don't exist. Twice.
  3. The 8GB VRAM ceiling. I can't run the 32B coder model without swapping. If you have 16GB+ VRAM this whole article gets more optimistic.

The Verdict

Acceptance rate dropped 7 points, but latency improved 4x and cost dropped to zero. For my mix (mostly Python/TS business logic, tests, boilerplate) local is a net win. For heavy greenfield work against unfamiliar APIs, I'd keep a cloud tool on standby.

The uncomfortable conclusion nobody in the "local AI" echo chamber says out loud: a 7B local model is 2024-Copilot, not 2026-Copilot. It's good enough that you stop paying, not good enough that you stop noticing.

I did the initial config and the acceptance-rate logging script with MonkeyCode (free tier) before moving fully offline — handy for the scaffolding phase: https://ly.cyberserval.tech/iIETXiF

Would you trade a 7-point acceptance drop for zero data leaving your machine? Or is the quality gap still a dealbreaker for you?

Top comments (0)