TopolearnModel choice drives quality, speed and cost, and different agents often need different models.
Every agent talks to an LLM. CrewAI names models as `provider/model-id`, for example openai/gpt-4o-mini or anthropic/claude-sonnet-4-6. The provider prefix is required. Model names change often, so check your provider's current list.
There are three places to choose a model:
- The
MODELenvironment variable (often in.env) sets the default for agents that don't specify one. - A string on the agent:
Agent(..., llm="anthropic/claude-sonnet-4-6"). - An
LLMobject, when you need settings:
from crewai import LLM
llm = LLM(model="openai/gpt-4o-mini", temperature=0.2, max_tokens=800)Common LLM parameters: temperature, max_tokens, timeout, api_key, base_url (a different server address) and stream.
OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock and a few others have native integrations, installed with extras like crewai[anthropic]. Everything else goes through LiteLLM, installed with crewai[litellm].
To run models on your own computer, use Ollama, a free app that serves open models locally: LLM(model="ollama/llama3.2", base_url="http://localhost:11434"). Local models are private and free per call, but smaller ones follow complex agent instructions and tool formats less reliably.
A practical pattern is mixing models: a strong model for planning or final writing, a cheaper, faster one for simple extraction steps. Some reasoning models reject temperature; if a call fails with a parameter error, remove it.
Example
from crewai import LLM, Agent
fast = LLM(model="openai/gpt-4o-mini", temperature=0.2) # needs OPENAI_API_KEY
strong = LLM(model="anthropic/claude-sonnet-4-6", max_tokens=2000) # needs crewai[anthropic]
local = LLM(model="ollama/llama3.2", base_url="http://localhost:11434") # crewai[litellm]
extractor = Agent(role="Data extractor", goal="Pull fields from text accurately",
backstory="You never guess missing values.", llm=fast)
editor = Agent(role="Editor", goal="Polish the final report",
backstory="You edit for clarity and accuracy.", llm=strong)
offline = Agent(role="Note taker", goal="Summarize private notes",
backstory="You work fully offline.", llm=local)
print(fast.model) # gpt-4o-miniRemember
- Models are named
provider/model-id; the prefix is required - Set a default with
MODEL, override per agent with a string orLLM(...) - Native providers via extras; others via
crewai[litellm]; local via Ollama
Watch out for
- Writing just "gpt-4o" without the provider prefix, which CrewAI cannot route.