Skip to content
Tutorials Center
My learning

Topics

Lesson 3 of 53 ยท Getting Started

The Model Landscape

Compare closed and open-weight language models, model sizes, context windows, multimodal and reasoning models, licensing, and learn how to choose one.

New language models appear every few weeks, and names and rankings go out of date quickly. What stays useful is the shape of the landscape: the kinds of models that exist, the trade-offs between them, and a repeatable way to pick one for your task. This lesson avoids product names on purpose. Instead it gives you the categories and the questions to ask, so the advice still works next year.

Closed versus open-weight

A closed (or proprietary) model is available only through a vendor's API or app. You send text, you get text back, and you never see the weights. The vendor runs the hardware, updates the model, and sets the price and rules.

An open-weight model publishes its trained weights so you can download and run them yourself, on your laptop, your servers or a rented GPU. "Open-weight" is more precise than "open source": you usually get the weights and often some code, but not always the training data or the full training recipe.

Closed or open-weight?

Closed API Easy start, top quality, pay per use.
Open weights You host it and control the data.

Many teams use both, for different jobs.

1 / 2
Question Closed API model Open-weight model
Setup effort An API key Hardware, runtime, serving
Data leaves your network Yes (check the vendor terms) No, if you self-host
Cost shape Pay per token Fixed hardware cost, cheap per token at scale
Customizing Limited fine-tuning options Full fine-tuning, quantization, merging
Model changes under you Vendor may retire or update versions You control the version
Peak quality Often the strongest Close behind, depends on the task

Model sizes

Size is counted in parameters: 1B means one billion. Rough tiers help with intuition:

  • Small (under about 10B): run on a laptop or phone, fast and cheap, good for classification, extraction and simple chat.
  • Medium (about 10B to 100B): need a serious GPU, noticeably better at reasoning and following long instructions.
  • Large (100B and above): usually served from data centres, strongest general ability, highest cost.

Bigger is usually better for the same recipe, but a small, well-trained model can beat an older large one. Some large models use a mixture of experts: only a part of the network is active for each token, so the model has many total parameters but costs less per token than a dense model of the same size.

Context windows

The context window is how many tokens the model can consider at once, covering your prompt, earlier conversation and its own answer. Windows range from a few thousand tokens to over a million. A longer window lets you paste whole documents, but costs more and models can still miss details buried in the middle (see Context Windows and the KV Cache). Always leave room for the answer: if the window is 8,000 tokens and your prompt is 7,900, only 100 tokens remain.

Multimodal models

A multimodal model accepts more than text: images, audio, sometimes video or documents with layout. Typical uses are reading a chart or screenshot, transcribing speech and describing photos. Some also produce images or audio. If your task is "extract the total from this invoice photo", you need vision input; if it is only text, a text-only model is simpler and cheaper. See Audio, Video and Multimodal Models.

Reasoning models

A reasoning model is trained to spend extra tokens "thinking" before it answers, working through a problem step by step in a hidden or visible scratchpad. They are often much better at math, logic, planning and multi-step code, and they are slower and more expensive per answer because of the extra tokens. For simple lookups, rewriting and classification a regular model is usually faster and just as good. Reasoning: Chain of Thought and Beyond explains how this connects to prompting.

Licensing

Check the license before you build on a model:

  • Some open-weight licenses are permissive and allow commercial use.
  • Others limit use above a user count, forbid certain uses, or require attribution or an acceptable-use policy.
  • Closed models are governed by the vendor's terms: look for rules about data retention, training on your inputs, and output ownership.
  • Outputs can also raise copyright and privacy questions; see Bias, Privacy, Copyright and Responsible Use.

This tutorial deliberately names no specific models. Capabilities, prices and licenses change often, so read the vendor's current documentation when you decide.

How to choose

Do not start from a leaderboard. Start from your requirements, filter, then test.

  1. Hard constraints first: data privacy, region, license, required modalities, needed context length, latency target.
  2. Shortlist two or three candidates of different sizes.
  3. Test on your own examples with an evaluation set (see Evaluating LLM Applications).
  4. Compare cost and speed at your real traffic, not on one demo prompt.
  5. Plan for change: keep model names in configuration so you can swap them.
Python A weighted scoring helper

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The helper shows two ideas worth remembering: the best model changes with the weights, and a hard requirement should exclude options outright rather than be traded off against price.

Next, a quick check that a prompt will fit a model's window and that the weights fit your hardware.

Python Does it fit? Context and memory

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Weights are only part of the memory story: the working memory for long prompts is covered in Context Windows and the KV Cache, and shrinking the weights with 8-bit or 4-bit numbers in LoRA, Quantization and Efficient Models.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What does open-weight mean?

    Open-weight means the weights are published. Training data and full recipes may still be private, and a license still applies.
  2. 2. When is a reasoning model most worth its extra cost and delay?

    Reasoning models spend extra tokens working through a problem, which pays off on hard multi-step tasks but is wasteful for simple ones.
  3. 3. A prompt uses 7,900 tokens and the model window is 8,000 tokens. What is the risk?

    The window covers the prompt and the response together, so a nearly full prompt leaves almost no space to answer.
  4. 4. What is the best first step when choosing a model for a product?

    Constraints such as privacy, modality and latency remove options, and your own test set shows which remaining model works for your task.

Summary

  • Closed models are easy to start with and often strongest; open-weight models give control, privacy and customization at the cost of running them.
  • Parameter count, context window, modalities and reasoning ability are the main axes that separate models.
  • Reasoning models trade speed and cost for better multi-step problem solving.
  • Always read the license and data terms; they decide what you may build.
  • Choose by constraints, shortlist, test on your own data, and keep model names configurable.

Next: send your first request in Calling an LLM API.