Skip to main content

Ollama

Serve open-weight language models on hardware you control — a laptop, a workstation, a small on-premise server — with an assistant that never sends a single document to the outside.

Course Duration: 4h​

What You'll Learn​

  • Install and operate the Ollama runtime on Windows, macOS and Linux
  • Pull, list and remove models, and pick a quantization tag that fits the machine
  • Drive an interactive session and set generation parameters that actually matter
  • Call the local API from Python, either natively or through the OpenAI-compatible endpoint
  • Author a Modelfile that pins system instructions, defaults and imported GGUF weights
  • Read RAM and VRAM requirements from a model's tag and diagnose out-of-memory failures
  • Accelerate inference on NVIDIA, AMD and Apple Silicon, and measure tokens per second
  • Integrate Ollama with LangChain, Open WebUI and a business script safely
  • Build a fully offline document question-answering system with local embeddings
  • Know when local execution is the right answer, and when to reach for a hosted API

Prerequisites​

  • Comfortable at the command line on your operating system
  • Course 16 — Large language models, recommended

Course Modules​

  1. Installing on Windows, macOS and Linux
  2. Pulling, listing and removing models
  3. Interactive session and generation parameters
  4. The local API
  5. Modelfiles and derived models
  6. Quantization and memory footprint
  7. Hardware acceleration
  8. Integrating with existing applications
  9. Local document question answering
  10. Limits of local execution

The running project​

Across the ten modules you build the same object: a private assistant for a small law firm that reviews contracts and answers questions about client files. Nothing leaves the office — not a name, not a clause, not a case number. Each module adds one capability to that same assistant: install (1), the right model in the client's working language (2), a stable session (3), a Python driver (4), a Modelfile that carries the firm's house style (5), enough quantization to fit a partner's laptop (6), a GPU that halves latency (7), a bridge to the existing case-management tool (8), a retrieval layer over the firm's PDF archive (9), and an honest discussion of what a local model still cannot do (10).

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.