Open vs closed models is a fundamental choice in AI solutions. Let us cover: 1) CLOSED models (proprietary, via API): — What it is: the model belongs to the provider, you use it via an API (send a request over the internet, get an answer). You do not get the model itself (its "weights"). Many leading powerful models are closed. — Pros: • Often the most powerful and advanced. • Convenience: no need to deploy anything, just call the API (from the LLM-API course). • The provider handles infrastructure, updates, scaling. • A fast start. — Cons: • Data goes to the provider: your requests (and the data in them) are sent out. A problem for private/sensitive data. • Pay per call: you pay for each request (by tokens). Costly at volume. • No control: you cannot change the model, you depend on the provider (changes, availability, prices, terms). • Dependence on the internet and an external service. 2) OPEN models (open-weight/open-source): — What it is: the model weights are published, you can DOWNLOAD and run it yourself — locally (on your computer) or on your server. There are many quality open models. — Pros: • PRIVACY: data does not leave — the model works with you, requests stay local. A huge plus for sensitive data. • CONTROL: the model is yours — you can tune, fine-tune, use it as you wish, without depending on a provider. • Cost at volume: no pay per call (you pay for your hardware/hosting). At large volume it can be cheaper than an API. • Autonomy: works without an external service/internet (after setup). • Transparency: more understanding and control over what the model is. — Cons: • RESOURCES needed: running the model requires hardware (memory, often a GPU for large models). Powerful models are demanding. • Setup and maintenance: you need to deploy, update, maintain it yourself (no provider doing it for you). • Often (not always) inferior to the most advanced closed ones in "raw" quality, though the gap is narrowing, and for many tasks open ones suffice. • More technical expertise required. 3) When to choose which: — Closed APIs are convenient when: you need a fast start, maximum quality, no sensitive data, moderate volume, do not want to deal with infrastructure. — Open/local models win when: • PRIVACY is critical (sensitive, personal, regulated data that cannot be sent out) — the main reason. • CONTROL is needed (customization, independence from a provider, stability of terms). • LARGE VOLUME where pay-per-call makes the API costly (your hardware pays off). • Autonomy (working without the internet/external services). • The task does not require the most advanced model — an open one suffices. — A hybrid: often combined — open for the sensitive/high-volume, closed for the complex/rare. Practical rule: closed models (API) — convenience and maximum quality, but data leaves and you pay per call; open/local — privacy, control and possible saving at volume, but resources, setup and expertise are needed. Choose for the requirements: privacy/control/volume to open local; a fast start/maximum quality/convenience to closed APIs. A hybrid is often optimal. This course is about open and local models: how to run them and when they give an advantage. Many principles of previous courses (LLM API, RAG, fine-tuning, LLMOps) apply to open models too — what mainly changes is WHERE the model runs (with you vs with the provider) and the related privacy, cost, control.