The developer — « Local Copilot, 100% offline, 0 cloud latency. »
Qwen2.5 Coder 32B connected to VSCode through Continue.dev. Autocomplete, chat, refactoring across an entire file — without a single line of your code leaking to Microsoft, OpenAI, or Anthropic. Does your PR contain an API key? It stays there.
At the end of this process, you'll have a llama.cpp server running in the background with Qwen2.5 Coder 32B in Q4 quantization. Continue.dev will be connected to VSCode with autocomplete (FIM) and sidebar chat. Latency will be under 100 ms, and autocomplete quality will be comparable to Copilot, sometimes better on recent code.
Who it's for this path is for you
We'd rather tell you early than let you waste 30 minutes for nothing.
- ✓You code every day and are tired of copying your snippets into a web UI.
- ✓Your employer bans GitHub Copilot, ChatGPT, and Cursor — for good reasons.
- ✓You're working with proprietary code, under an NDA, or with API secrets in plain text.
- ✓You have a RTX 4070 Ti+, RTX 4090, RTX 5090, or a MacBook Pro M2/M3/M4 with 32 GB+.
- ✓The 200 ms ping to cloud APIs breaks your rhythm.
- ·You have 8 GB of VRAM: it's workable, but with a 7B model, quality is well below Copilot.
- ·You code 1h a week: a free account with a cloud provider will be less of a hassle.
- ·You want general assistant chat: see the Curious path, which is simpler.
The path in 4 steps
Each step points to a detailed guide you can read alongside it. The order is optimized: don’t skip a step the first time.
- 1Step 01·12 min·Advanced
Compile llama.cpp with GPU acceleration
Build from source for maximum tokens/sec. On NVIDIA: CUDA. On Apple Silicon: Metal. On AMD: ROCm. It’s 15 to 30% faster than Ollama.
Ready-to-use llama-server binary capable of serving an OpenAI-compatible API.Read the guide → - 2Step 02·6 min·Intermediate
Download Qwen2.5 Coder 32B (Q4)
The best open-weight coding model in its class. Q4_K_M fits in ~19 GB of VRAM while retaining 95% of FP16 quality. Download from Hugging Face.
The model is served on localhost:8080 in full chat completions format.Read the guide → - 3Step 03·8 min·Intermediate
Connect Continue.dev to VSCode
Install the Continue extension from the marketplace, edit ~/.continue/config.json to point to your local endpoint, and configure the FIM model for autocompletion.
Tab to accept a suggestion, Cmd+L to open chat for the selection.Read the guide → - +Step bonus·4 min·Advanced
Bonus — Helping from the command line
For larger tasks (refactoring a module, generating tests), Aider lets you edit files in natural language from the terminal, with automatic Git commits.
An agentic CLI connected to the same backend, complementary to Continue.Read the guide →
Recommended models
Three choices tested for this path—from lightest to most capable. Click for the detailed specifications (VRAM, context, benchmarks).
The versatile one. 14B = 9 GB in Q4, excellent code quality, fits on 12 GB of VRAM.
A good compromise if you have 16 GB VRAM, with broader capabilities.
View details →The best choice if you have a 64 GB Mac or a 48 GB GPU. Higher latency but top-tier quality.
Frequently asked questions
The real questions we get by email and on Mastodon. If yours is missing, open a ticket.
The confidential pro — « My files never leave my workstation. »
Juriste, médecin, RH, consultant, expert-comptable, notaire — vos documents sont sous secret professionnel ou NDA. Un LLM local couplé à un RAG vous laisse les interroger, résumer,…
Replace GitHub Copilot and Cursor with a code assistant that runs 100% on your machine — the reference guide, configs included.
- Lifetime online access
- PDF + files
- Lifetime updates
A question, a typo, a bug?
This guide evolves with every model release. Your feedback is the raw material.