Choosing a model

A model is a trained neural network, named by an id such as openai/gpt-oss-120b, and choosing one means trading answer quality against speed and price.

Last updated: 30 Sep, 2026 · groq 1.7 · gpt-oss-120b on Groq

With a hosted API you never load a model yourself: every request names the model it wants, and the provider runs it. Changing models is changing one string.

The video describes a language model by two numbers. Parameters are its weights and biases, the numbers it learned, counted in billions or trillions. Tokens are the pieces of text it was trained on, also counted in trillions. The transformer is the architecture underneath. The video then uses parameters to sort models: below about 10 billion it calls a model a small language model (SLM).

That 10 billion line is a rule of thumb from the video, not an official definition. The two models below carry their size in their names: gpt-oss-20b has about 20 billion parameters and gpt-oss-120b about 120 billion.

Syntax: client.models.list() returns the models your key can use; model= in a request picks one.

Listing the models your key can use

from groq import Groq

client = Groq()  # reads GROQ_API_KEY from the environment

for model in client.models.list().data:
    print(model.id)

whisper-large-v3
openai/gpt-oss-safeguard-20b
canopylabs/orpheus-v1-english
meta-llama/llama-prompt-guard-2-22m
canopylabs/orpheus-arabic-saudi
qwen/qwen3.8-27b
openai/gpt-oss-20b
allam-2-7b
openai/gpt-oss-120b
meta-llama/llama-prompt-guard-2-86m
whisper-large-v3-turbo

• Not every id is a chat model. The whisper models turn speech into text, the orpheus models turn text into speech, and the prompt-guard models only classify prompts.
• The chat models include openai/gpt-oss-20b and openai/gpt-oss-120b, the pair this lesson compares.
• The order changes from call to call; sort the ids if you want a stable list.

Details of the two gpt-oss models

models = {model.id: model for model in client.models.list().data}

for name in ["openai/gpt-oss-20b", "openai/gpt-oss-120b"]:
    model = models[name]
    print(name, "| made by", model.owned_by, "| weights:", model.hugging_face_id)

openai/gpt-oss-20b | made by OpenAI | weights: openai/gpt-oss-20b
openai/gpt-oss-120b | made by OpenAI | weights: openai/gpt-oss-120b

Both are made by OpenAI, and their weights are published on Hugging Face under the same names. That is what open-weight means: you could download them and run them yourself, on hardware with enough memory. Groq runs them for you.

Asking the 20b and the 120b the same question

question = [{"role": "user", "content": "In one sentence, what is a refund?"}]

for name in ["openai/gpt-oss-20b", "openai/gpt-oss-120b"]:
    response = client.chat.completions.create(model=name, messages=question)
    print(name)
    print("  ", response.choices[0].message.content)

openai/gpt-oss-20b
   A refund is the return of money paid for a product or service, usually given to a buyer who has returned or is dissatisfied with the purchase.
openai/gpt-oss-120b
   A refund is the return of money to a buyer when a purchased product or service is returned, canceled, or found unsatisfactory.

What the two answers show

• Both answers are correct and one sentence long, as asked.
• The wording differs. The 20b answer mentions a buyer who is dissatisfied; the 120b answer lists three cases: returned, cancelled or unsatisfactory.
• An easy question hides the difference in size. Larger models pull ahead on harder tasks: long instructions, tricky categories, several steps of reasoning. Project: ticket sorter swaps models on a real task and measures it.

Hosted model vs open weights on your computer

When to pick the smaller model

• A simple, well-defined task, such as sorting tickets into three categories, where the smaller model's score is good enough.
• High volume, where half the price per token adds up (Cost reads both prices).
• When speed matters more than polish.

Related

• Previous: Installation and setup
• Next: Tokens
• Reference: Groq supported models