​
1. Concept

The model is the brain of the intelligent agent, responsible for intelligent agent, processing all input and output data to execute tasks such as text analysis, image recognition, and complex reasoning. With customizable interfaces and multiple integration options, CAMEL AI enables rapid development with leading LLMs.

Explore the Code: Check out our Colab Notebook for a hands-on demonstration.

​
2. Supported Model Platforms in CAMEL

CAMEL supports a wide range of models, including OpenAI’s GPT series, Meta’s Llama models, DeepSeek models (R1 and other variants), and more. The table below lists all supported model platforms:

Model PlatformModel Type(s)
OpenAIgpt-4.5-preview, gpt-4o, gpt-4o-mini, o1, o1-preview, o1-mini, o3-mini, o3-pro, gpt-4-turbo, gpt-4, gpt-3.5-turbo
Azure OpenAIgpt-4o, gpt-4-turbo, gpt-4, gpt-3.5-turbo
Mistral AImistral-large-latest, pixtral-12b-2409, ministral-8b-latest, ministral-3b-latest, open-mistral-nemo, codestral-latest, open-mistral-7b, open-mixtral-8x7b, open-mixtral-8x22b, open-codestral-mamba, magistral-medium-2506
Moonshotmoonshot-v1-8k, moonshot-v1-32k, moonshot-v1-128k
Anthropicclaude-2.1, claude-2.0, claude-instant-1.2, claude-3-opus-latest, claude-3-sonnet-20240229, claude-3-haiku-20240307, claude-3-5-sonnet-latest, claude-3-5-haiku-latest
Geminigemini-2.5-flash-preview-04-17, gemini-2.5-pro-preview-06-05, gemini-2.0-flash, gemini-2.0-flash-exp, gemini-2.0-flash-thinking-exp, gemini-2.0-pro-exp-02-05, gemini-2.0-flash-lite, gemini-2.0-flash-lite-preview-02-05, gemini-1.5-flash, gemini-1.5-pro
Lingyiwanwuyi-lightning, yi-large, yi-medium, yi-large-turbo, yi-vision, yi-medium-200k, yi-spark, yi-large-rag, yi-large-fc
Qwenqwq-32b-preview, qwen-max, qwen-plus, qwen-turbo, qwen-long, qwen-vl-max, qwen-vl-plus, qwen-math-plus, qwen-math-turbo, qwen-coder-turbo, qwen2.5-coder-32b-instruct, qwen2.5-72b-instruct, qwen2.5-32b-instruct, qwen2.5-14b-instruct
DeepSeekdeepseek-chat, deepseek-reasoner
ZhipuAIglm-4, glm-4v, glm-4v-flash, glm-4v-plus-0111, glm-4-plus, glm-4-air, glm-4-air-0111, glm-4-airx, glm-4-long, glm-4-flashx, glm-zero-preview, glm-4-flash, glm-3-turbo
InternLMinternlm3-latest, internlm3-8b-instruct, internlm2.5-latest, internlm2-pro-chat
Rekareka-core, reka-flash, reka-edge
COHEREcommand-r-plus, command-r, command-light, command, command-nightly
GROQsupported models
TOGETHER AIsupported models
SambaNovasupported models
Ollamasupported models
OpenRoutersupported models
PPIOsupported models
LiteLLMsupported models
LMStudiosupported models
vLLMsupported models
SGLANGsupported models
NetMindsupported models
NOVITAsupported models
NVIDIAsupported models
AIMLsupported models
ModelScopesupported models
AWS Bedrocksupported models
IBM WatsonXsupported models
Crynuxsupported models

​
3. How to Use Models via API Calls

Easily integrate your chosen model with CAMEL AI using straightforward API calls. For example, to use the gpt-4o-mini model:

If you want to use another model, you can simply change these three parameters: model_platform, model_type, model_config_dict .

from camel.models import ModelFactory
from camel.types import ModelPlatformType, ModelType
from camel.configs import ChatGPTConfig
from camel.messages import BaseMessage
from camel.agents import ChatAgent

# Define the model, here in this case we use gpt-4o-mini
model = ModelFactory.create(
    model_platform=ModelPlatformType.OPENAI,
    model_type=ModelType.GPT_4O_MINI,
    model_config_dict=ChatGPTConfig().as_dict(),
)

# Define an assistant message
system_msg = "You are a helpful assistant."

# Initialize the agent
ChatAgent(system_msg, model=model)

And if you want to use an OpenAI-compatible API, you can replace the model with the following code:

model = ModelFactory.create(
    model_platform=ModelPlatformType.OPENAI_COMPATIBLE_MODEL,
    model_type="a-string-representing-the-model-type",
    api_key=os.environ.get("OPENAI_COMPATIBILITY_API_KEY"),
    url=os.environ.get("OPENAI_COMPATIBILITY_API_BASE_URL"),
    model_config_dict={"temperature": 0.4, "max_tokens": 4096},
)

​
4. Using On-Device Open Source Models

CAMEL AI also supports local deployment of open-source LLMs. Choose the setup that suits your project:

​
4.1 Using Ollama to Set Llama 3 Locally

  1. Download Ollama.
  2. After setting up Ollama, pick a model like Llama3 for your project:
ollama pull llama3
  1. Create a ModelFile similar the one below in your project directory. (Optional)
FROM llama3

# Set parameters
PARAMETER temperature 0.8
PARAMETER stop Result

# Sets a custom system message to specify the behavior of the chat assistant
# Leaving it blank for now.

SYSTEM """ """
  1. Create a script to get the base model (llama3) and create a custom model using the ModelFile above. Save this as a .sh file: (Optional)
#!/bin/zsh

# variables
model_name="llama3"
custom_model_name="camel-llama3"

#get the base model
ollama pull $model_name

#create the model file
ollama create $custom_model_name -f ./Llama3ModelFile
  1. Navigate to the directory where the script and ModelFile are located and run the script. Enjoy your Llama3 model, enhanced by CAMEL’s excellent agents.
from camel.agents import ChatAgent
from camel.messages import BaseMessage
from camel.models import ModelFactory
from camel.types import ModelPlatformType

ollama_model = ModelFactory.create(
    model_platform=ModelPlatformType.OLLAMA,
    model_type="llama3",
    url="http://localhost:11434/v1", # Optional
    model_config_dict={"temperature": 0.4},
)

agent_sys_msg = "You are a helpful assistant."

agent = ChatAgent(agent_sys_msg, model=ollama_model, token_limit=4096)

user_msg = "Say hi to CAMEL"

assistant_response = agent.step(user_msg)
print(assistant_response.msg.content)

​
4.2 Using vLLM to Set Phi-3 Locally

Install vLLM first.

After setting up vLLM, start an OpenAI compatible server for example by:

python -m vllm.entrypoints.openai.api_server --model microsoft/Phi-3-mini-4k-instruct --api-key vllm --dtype bfloat16

Create and run following script (more details please refer to this example):

from camel.agents import ChatAgent
from camel.messages import BaseMessage
from camel.models import ModelFactory
from camel.types import ModelPlatformType

vllm_model = ModelFactory.create(
    model_platform=ModelPlatformType.VLLM,
    model_type="microsoft/Phi-3-mini-4k-instruct",
    url="http://localhost:8000/v1", # Optional
    model_config_dict={"temperature": 0.0}, # Optional
)

agent_sys_msg = "You are a helpful assistant."

agent = ChatAgent(agent_sys_msg, model=vllm_model, token_limit=4096)

user_msg = "Say hi to CAMEL AI"

assistant_response = agent.step(user_msg)
print(assistant_response.msg.content)

​
4.3 Using SGLang to Set meta-llama/Llama Locally

Install SGLang first.

Create and run following script (more details please refer to this example):

from camel.agents import ChatAgent
from camel.messages import BaseMessage
from camel.models import ModelFactory
from camel.types import ModelPlatformType

sglang_model = ModelFactory.create(
    model_platform=ModelPlatformType.SGLANG,
    model_type="meta-llama/Llama-3.2-1B-Instruct",
    model_config_dict={"temperature": 0.0},
    api_key="sglang",
)

agent_sys_msg = "You are a helpful assistant."

agent = ChatAgent(agent_sys_msg, model=sglang_model, token_limit=4096)

user_msg = "Say hi to CAMEL AI"

assistant_response = agent.step(user_msg)
print(assistant_response.msg.content)

​
5 Model Speed and Performance

Performance is critical for interactive AI applications. CAMEL-AI benchmarks tokens processed per second across various models:

In this notebook, we compared several models, including OpenAI’s GPT-4O Mini, GPT-4O, O1 Preview, and SambaNova’s Llama series, by measuring the number of tokens each model processes per second.

Key Insights: Smaller models like SambaNova’s Llama 8B and OpenAI’s GPT-4O Mini typically offer faster responses. Larger models like SambaNova’s Llama 405B, while more powerful, tend to generate output more slowly due to their complexity. OpenAI models demonstrate relatively consistent performance, while SambaNova’s Llama 8B significantly outperforms others in speed. The chart below illustrates the tokens per second achieved by each model during our tests:

For local inference, we conducted a straightforward comparison locally between vLLM and SGLang. SGLang demonstrated superior performance, with meta-llama/Llama-3.2-1B-Instruct reaching a peak speed of 220.98 tokens per second, compared to vLLM, which capped at 107.2 tokens per second.

​
6. Next Steps

You’ve now learned how to integrate various models into CAMEL AI.

Next, check out our guide covering basics of creating and converting messages with BaseMessage.