Skip to main content
This integration connects Ollama’s local models to LangChain.

​
Installation

First, install Ollama from ollama.com and pull a model: 
ollama pull llama3.2
Then install the LangChain integration: 
pip install -U langchain-ollama

​
Usage

from langchain_ollama import ChatOllama

model = ChatOllama(
    model="llama3.2",
    temperature=0.8,
    num_predict=256,
)

messages = [
    ("system", "You are a helpful assistant."),
    ("human", "What is the capital of France?"),
]

response = model.invoke(messages)
print(response.content)

​
Streaming

for chunk in model.stream(messages):
    print(chunk.content, end="")

​
API Reference

​
ChatOllama

​
model
str
required
Name of Ollama model to use (e.g., llama3.2, mistral, phi3).
​
temperature
float
default:"0.8"
Sampling temperature between 0.0 and 1.0. Higher values make output more random.
​
num_predict
int | None
default:"None"
Maximum number of tokens to generate.
​
reasoning
bool | None
default:"None"
Controls reasoning/thinking mode for supported models:
  • True: Enables reasoning mode. Reasoning is captured in additional_kwargs.reasoning_content
  • False: Disables reasoning mode
  • None: Uses model’s default behavior
​
base_url
str
default:"http://localhost:11434"
Base URL where Ollama is running.
​
top_k
int | None
default:"None"
Reduces probability of generating nonsense. Higher values give more diversity.
​
top_p
float | None
default:"None"
Works together with top_k. Higher values give more diversity.
​
num_ctx
int | None
default:"None"
Sets the size of the context window used to generate the next token.
​
repeat_penalty
float | None
default:"None"
Sets how strongly to penalize repetitions. Higher values make repetitions less likely.
​
validate_model_on_init
bool
default:"False"
Whether to validate that the model exists when initializing.
​
stop
list[str] | None
default:"None"
Stop sequences to end generation.

​
Supported Models

Ollama supports hundreds of models. Popular options include:
  • Llama 3.2: Fast, efficient model from Meta
  • Mistral: High-quality open model
  • Phi-3: Microsoft’s small language model
  • Gemma: Google’s open model
  • DeepSeek: Reasoning-capable models
  • Qwen: Alibaba’s multilingual models
Visit ollama.com/library for the full model catalog.

​
Features

  • Run models locally without API keys
  • Full privacy - no data sent to external servers
  • Tool calling (select models)
  • Vision capabilities (multimodal models)
  • Streaming
  • Async support
  • Custom model parameters
  • Reasoning mode for supported models
Ollama runs models locally on your machine. Performance depends on your hardware. GPU acceleration is recommended for larger models.

Build docs developers (and LLMs) love

Get started for freeTalk to us