A field guide, not a fan page.
Inferwerks teaches the two disciplines that turn a model into a product: inference engineering and harness engineering. It is free, static, and built to be read slowly.
Why this exists
Most material about LLM systems is either a vendor benchmark or a prompt tutorial. The interesting work sits between them: why a request waits, why a cache fills, why an agent loops, why a cheap model becomes an expensive incident.
The site treats inference and harnesses as one connected system. A serving SLO is part of the product experience. An agent’s tool call is an inference request. The failure is usually in the seam.
The two themes
Draft is a warm paper notebook: serif display type, ink lines, cobalt and orange, drafting-table diagrams. Live is a night signals desk: indigo surfaces, electric cyan, hot magenta and violet traces, grotesque display type.
They are the same information in two visual languages, not a light-mode afterthought. The toggle is in the header. Your choice is stored in this browser only.
Editorial rules
- Explain the mechanism before naming the optimization.
- Label estimates as estimates. Date the figures. Link the source.
- Do not present a benchmark as a capacity plan.
- Distinguish a claim from evidence. “The model says done” is not verification.
- Prefer a small, testable model over a large, unread one.
- When a topic is contested or moving quickly, say so.
How it is built
Astro and MDX generate a fully static site deployed on Cloudflare Pages. Preact islands power the three labs. KaTeX renders the math. Pagefind indexes the built pages for search. There is no backend, no account system and no analytics.
Progress is stored in localStorage under a single key. Clearing site data removes it. The workbenches run entirely in your browser; nothing you type is sent anywhere.
Accuracy and corrections
Hardware figures come from vendor documentation. Papers are linked on the Sources page. The calculators use simplified analytical models; they are not benchmarks. If you find an error, open an issue in the repository with the lesson, the claim and the source that contradicts it.
The curriculum is dated November 2026. Serving engines and model families change quickly; verify current behavior before making an adoption decision.
License and use
Content is licensed under CC BY-NC-SA 4.0. Code is MIT. You may read, link and teach with it. If you reuse a diagram, credit the site and link back. Do not republish the curriculum as a paid course.
The site is independent. It is not affiliated with any model provider, GPU vendor or agent harness mentioned in the lessons.
The model generates. The harness decides.
Start with the primer, or jump straight to the track you need. The map is always there when you get lost.