A cookbook that treats open models as a workflow, not a toy

What it does

The ModelScope Cookbook (魔搭紫皮书) is a structured, hands-on book about building real applications with open-source AI models. It covers the full arc — model selection, inference, data prep, fine-tuning, evaluation, RAG, agents, and generative image/video work — tied together with runnable examples built on tools like EvalScope, ms-swift, DiffSynth, and Ollama. The framing is refreshingly practical: which model fits the task, what hardware it needs, how to adapt it, and how to judge whether the result is any good.

The interesting bit

Instead of chapters about concepts, most chapters are end-to-end mini-projects: an AI fitness coach that compares body keypoints, a customer-service quality analyzer that digests call recordings, a phone-based factory inspection assistant, a product-image generator. The learning-paths table maps goals (get started, adapt a model, build an app, do AIGC) to specific chapter sequences, so you can skip the linear reading.

Key highlights

• 8 parts, 38 chapters, 37 currently readable — one is explicitly a placeholder pending a MiniMax H3 tutorial
• Reading paths for four distinct goals, from first inference to agent development
• Covers the unglamorous middle: quantization trade-offs, server sizing, turning business materials into training data
• Readable online with full-text search, or locally with nothing but Python 3 and a browser — no build step, no API key, no GPU for the reading itself
• Apache 2.0, contributions welcome, and chapter 22 invites readers to add their own reproducible scenarios

Caveats

• Chapter titles and content are in Chinese; the README’s English translation only covers the navigation layer
• One chapter (27, AI video) is still pending, and chapter 38 is an open invitation rather than finished content
• Individual exercises do require model downloads, credentials, or compute — the zero-dependency claim applies to reading, not doing

Verdict

Worth a look if you want a curated, project-driven path through the open-model ecosystem and read Chinese (or don’t mind machine translation). If you wanted a finished English reference, it’s not there yet — but the structure and example choices are a good map regardless.