Skip to content

3. Model configuration

Jump to bottom
fanlv edited this page Oct 22, 2025 · 10 revisions

Coze Studio is an AI app development platform based on LLM. Before running the Coze Studio open-source version for the first time, you need to clone the project to your local machine and configure the required models. During normal project operations, you can also add new model services as needed or delete unnecessary model services at any time.

Model list

The model services supported by Coze Studio are as follows:

  • Volcengine Ark | BytePlus ModelArk
  • OpenAI
  • DeepSeek
  • Claude
  • Ollama
  • Qwen
  • Gemini

Model configuration instructions

Before filling out the model configuration file, ensure you have understood the following Important notes:

  • Before deleting the model, ensure that it is no longer receiving online traffic.
  • Agents or workflows call models based on model IDs. For models that have already been launched, do not modify their IDs; otherwise, it may result in model call failures.

Configure the model for Coze Studio.

Coze Studio is an AI app development platform based on LLMs. Before deploying and launching the open-source version of Coze Studio for the first time, you need to configure the model service in the Coze Studio project, otherwise, you won't be able to properly select a model during the creation of agents or workflows.

Configure the model at http://localhost:8888/admin/#model-management by adding a new model.

Third-Party Model Service

Platform Protocol Base_url Special Instructions
Volcengine Ark ark Volcengine Engine: https://ark.cn-beijing.volces.com/api/v3/
Overseas BytePlus: https://ark.ap-southeast.bytepluses.com/api/v3/ 		None
Alibaba Bai Lian openai or
qwen		 https://dashscope.aliyuncs.com/compatible-mode/v1 The qwen3 series does not support thinking in non-streaming calls. If used, you need to set enable_thinking: false in conn_config. Coze Studio will adapt this capability in future versions.
Silicon-based Flow openai https://api.siliconflow.cn/v1 None
Other Third-Party API Relay openai The address provided in the API document
Note that the path usually has a /v1 suffix and does not have a /chat/completions suffix		 If the platform only relays or proxies model services and the model is not an openai model, please configure the protocol according to the documentation in the [Official Model Service] section.

Open-Source Framework

Framework Protocol Base_url Special Instructions
Ollama ollama http://${ip}:11434 1. When the mirror network mode is bridge, localhost in the coze-server mirror is not the localhost of the host. It needs to be modified to the ip of the Ollama deployment machine or http://host.docker.internal:11434.
2. Check the api_key: When the api_key is not set, this parameter is left blank.
3. Confirm whether the firewall of the Ollama-deployed host has opened port 11434.
4. Confirm that the Ollama network has enabled External Exposure.

Image
vllm openai http://${ip}:8000/v1 (specified when starting the port) None
xinference openai http://${ip}:9997/v1 (specified when starting the port) None
sglang openai http://${ip}:35140/v1 (specified when starting the port) None
LMStudio openai http://${ip}:${port}/v1 None

Official Model Service

Model Protocol Base_url Special Instructions
Doubao ark https://ark.cn-beijing.volces.com/api/v3/ None
OpenAI openai https://api.openai.com/v1 Check the by_azure field configuration. If the model is a model service provided by Microsoft Azure, this parameter should be set to true.
Deepseek deepseek https://api.deepseek.com/ None
Qwen qwen https://dashscope.aliyuncs.com/compatible-mode/v1 The qwen3 series does not support thinking in non-streaming calls. If used, you need to set enable_thinking: false in conn_config. Coze Studio will adapt this capability in future versions.
Gemini gemini https://generativelanguage.googleapis.com/ None
Claude claude https://api.anthropic.com/v1/ None

Home

Clone this wiki locally