Ollama¶
Ollama is a tool for running large language models locally — download and run models like Llama, Mistral, Qwen, Gemma, and DeepSeek on your own machine with a single command. It's the most popular on-ramp to self-hosted inference, and the sovereignty answer to "do I have to send everything to OpenAI/Anthropic?"
Why it matters for owned AI infrastructure¶
Local models mean prompts and data never leave your hardware — the ultimate privacy stance, and it removes per-token API bills for suitable workloads. The tradeoff: local models are weaker than frontier hosted models and need real hardware (GPU/RAM) for good performance.
Key facts¶
- Single-command model pull/run:
ollama run llama3. - Bundles model weights, config, and a REST API (OpenAI-compatible endpoint) so apps/agents can point at it.
- Runs on macOS, Linux, Windows; GPU-accelerated where available, CPU fallback.
- Model library covers most popular open-weight families; supports custom
Modelfilebuilds and quantized variants. - Can be self-hosted via docker-compose and fronted by caddy like any service.
What it replaces¶
Per-token calls to hosted LLM APIs (OpenAI/Anthropic) for workloads a smaller local model can handle. Not a full frontier-model replacement — a complement.
Related¶
- ai-agent-orchestration — local models as swappable agent backends
- docker-compose — deployment
- rag-vs-wiki — retrieval strategies work with local or hosted models alike