← Back to Catalog
Agent Skill

local-models

Run quick, private LLM tasks offline via llama.cpp.

local-models

Run quick, offline, private LLM tasks on local models via llama.cpp, reusing models already downloaded by Ollama: summarize, classify, extract JSON, anonymize PII, translate, embed, and describe images.

What it does

Gives quick command-line access to small local LLMs through llama.cpp, reusing the GGUF models already pulled by Ollama — no re-download, no API key, no per-token cost. An lm wrapper exposes one-shot prompting plus task presets (summarize, TL;DR, keywords, proofread, anonymize, translate, classify, extract-to-JSON), local embeddings, offline image description, and an OpenAI-compatible server. Everything runs on the machine.

Key features

  • Reuses Ollama’s downloads — reads Ollama’s manifests to resolve a friendly name like qwen2.5:3b straight to its on-disk GGUF blob, so llama.cpp loads it directly with zero duplication.
  • Privacy-first presetsanonymize redacts names, emails, phones and addresses into [NAME]/[EMAIL] placeholders without the text ever leaving the machine; ideal for transcripts and personal notes.
  • Clean, scriptable output — drives llama-completion in single-turn mode and strips the chat-template scaffolding and engine logs, so presets emit just the answer (a label, JSON, or prose) ready to pipe.
  • Embeddings and local RAGlm embed returns OpenAI-style vectors from multilingual embedding models, with a minimal cosine-similarity retrieval pattern for offline search.
  • Offline vision — describes and tags images via a bundled vision model, auto-fetching the projector llama.cpp needs on first use, then running fully offline.
  • Serve when you need speedlm serve keeps a model resident behind an OpenAI-compatible endpoint for high-volume or repeated calls.

When to use

When a task is privacy-sensitive, must run offline, is high-volume and low-stakes, or just needs a fast throwaway answer — and a round-trip to a frontier cloud model would be overkill.