Local Model Advisor

Find local AI models that actually fit your machine — from inside VS Code.

One scan reads your CPU, RAM and GPU, cross-references the most-downloaded GGUF models on Hugging Face, and ranks what your hardware can really run:

• GPU TURBO — fits entirely in VRAM (fastest)
• HYBRID — partial GPU offload
• CPU OK — fits in system RAM

Each recommendation ships with a one-click ollama run hf.co/<model> command.

Usage

1. Click the Local Model Advisor icon in the activity bar (the left-hand strip). The view opens idle — nothing runs until you ask it to.
2. Hit SCAN. Re-opening the sidebar later replays the last results instead of re-scanning.
3. Work through the results:
   • Filter by tier — All / GPU TURBO / HYBRID / CPU OK.
   • Sort by Recommended, params (↓/↑), downloads, size, name, source site, company of origin, or model family.
   • Page through every model that fits, 25 at a time — there is no top-N cut.
   • Click a model name to open its Hugging Face page, or ⧉ ollama run to copy its command.

The command palette still works: Local Model Advisor: Scan Hardware & Recommend Models focuses the sidebar and starts a scan.

Company of origin and model family are read from each model's Hugging Face base_model: lineage, not from the repository owner — most popular GGUF repos belong to re-packagers rather than the lab that trained the model. When the lineage does not resolve, the company reads Unknown rather than a guess.

Optional: Hugging Face token

The advisor works without any account. If you hit API rate limits, add a free read token from https://huggingface.co/settings/tokens via Local Model Advisor: Set Hugging Face Token — it is stored in your OS keychain (VS Code SecretStorage), never in settings or on disk in plain text.

Privacy

Hardware details never leave your machine. The only network call is a public model-listing request to huggingface.co; when offline, a bundled catalog is used instead.

How sizes are estimated

Sizes assume Q4 quantization (~0.6 GB per billion parameters) plus context headroom. Reported VRAM can be inaccurate on some Windows drivers — treat tiers as guidance, not gospel.