Llama, Mistral and Qwen on hardware you control
For regulated data, predictable costs or plain independence, open-weight models are now genuinely viable. This track is the practical version: choose a model, serve it properly, fine-tune it on your data, and know the real cost per token before you commit.
Ordered the way we teach it. Private workshops reorder and reweight this against what your team already knows.
Licences, parameter counts and benchmark inflation — how to shortlist for your task rather than for a leaderboard.
Throughput, batching and GPU sizing with a real inference server, so one user's long prompt doesn't stall everyone else's.
Fitting a larger model on smaller hardware, and measuring the quality you gave up rather than hoping it was free.
Parameter-efficient training end to end: dataset construction, a training run, and an evaluation that says whether it helped.
A complete RAG pipeline where no document and no query ever leaves your network.
Self-hosted versus API, worked through with your actual volumes — including the engineering time nobody puts in the spreadsheet.
Teams often pair a model track with Agentic Systems — the tool design and evaluation material applies whichever provider you settle on.
Send over your stack, your group size and roughly when you'd like it. We'll come back with a curriculum outline and a date.