Learn to select, run and adapt open-weight LLMs on your own infrastructure. Use measured GPU results to compare quantization, fine-tuning and serving, then apply them to enterprise RAG, agents and evaluation. Course material is in English.
What's inside
Compare open-weight model families by quality, licence, language support, context and memory.
Run the same GGUF file with Ollama and llama.cpp and compare generation speed.
Estimate memory for weights and KV cache and measure how context length affects inference.
Compare memory, speed and task quality at different precisions using the same model.
Design prompts, structured responses and tool calls for local models.
Explore how to transfer behaviour to a smaller model and when to use supervised adaptation.
Train a QLoRA adapter on an RTX GPU and compare answers before training and after reloading the adapter.
Build a local RAG pipeline with embeddings, retrieval, reranking and citations.
Define versioned evaluation sets to compare models and adaptations on specific tasks.
Measure vLLM throughput and latency at different concurrency levels and monitor GPU usage.
Size a local solution and compare hardware purchase, GPU rental and per-token costs.
Examine model licences, model provenance and data protection for a local deployment.
Use OpenCode with Ollama and compare local models on the same coding task.
Examine the components of an enterprise agent architecture on private infrastructure.
Evaluate local models for classification and decision tasks, comparing calibration, cascades and human review.