Offline autocomplete for the GPU-privileged developer

What it does Llama Coder is a VS Code extension that offers inline code completion by talking to a local or remote Ollama instance. It pulls quantized models—mostly CodeLlama variants and Stable Code—and runs them on your own hardware so your keystrokes never leave the building. The README claims the experience is “as good as Copilot,” provided you have the RAM and GPU to feed it.

The interesting bit The project treats your editor as a thin client for Ollama, but adds just enough polish—pause controls, bearer-token remote auth, Jupyter notebook support, and trigger delays—to make local inference feel like a product rather than a science experiment. It is also refreshingly blunt about the hardware tax, openly recommending an RTX 4090 or a MacBook with buckets of RAM.

Key highlights

• No telemetry or cloud phoning; everything stays on your LAN or machine.
• Supports remote Ollama endpoints with Bearer token authentication.
• Offers a tiered model menu from a 3GB stable-code:3b-code-q4_0 up to a 32GB codellama:34b-code-q6_K, with explicit warnings about MacOS and older NVIDIA slowness.
• Includes quality-of-life features like pausing completions, custom temperature and top-p settings, and Jupyter notebook support.

Caveats

• The README contradicts itself on model support: one section claims “only Codellama” while the table and changelog list Stable Code and DeepSeek.
• Minimum 16 GB RAM is required, and the “best” experience demands high-end Apple Silicon or an RTX 4090—this is not a lightweight browser extension.
• Several larger or higher-precision models are explicitly flagged as slow on MacOS or pre-30xx NVIDIA cards.

Verdict Grab it if you want Copilot-style autocomplete inside VS Code but refuse to ship code context to Microsoft’s servers. Skip it if you’re on a memory-constrained machine or hoping for a plug-and-play experience that doesn’t involve managing Ollama endpoints and VRAM budgets.