Can Coze call local large models? Yes, but only if you are talking about Coze Studio as a deployable environment, rather than understanding the web as a regular chat site. There are many people in the community who ask this question, and the most common scenario is that you run LM Studio, Ollama or other model services locally and want to connect them directly to Coze.
Coze Studio official open-source repositories are https://github.com/coze-dev/coze-studio. The official README has positioned it as a one-stop AI Agent development tool, so it's not that the local model can't be connected, but you need to check whether the model interface, protocol, and configuration file are really aligned.
Don't rush to ask "can you do it", first ask "what is the interface"
Whether the local model can be connected or not depends not on whether the model is local, but whether it provides interfaces that Coze can recognize. Many people think that as long as one port is opened, it actually depends on it:
- Is it compatible with OpenAI-style interfaces?
- Is there a correct 'base_url'?
- Are the model names, protocol types, and capability declarations consistent?
The most common stuck points in the community
The first category refers to the page address instead of base_url the API address. The second type is that the model service can chat, but does not support the request method that Coze needs. The third type is that the local model itself is fine, but you choose the wrong protocol in Coze, such as writing a model that should go to the OpenAI compatibility layer as another protocol.
Another point that is easy to overlook is that whether a model can "run natively" does not mean whether it can be "tuned up by Coze". The former is a model layer problem, and the latter is a request layer problem, which often needs to be viewed separately.
A relatively stable connection
If your local model provides an OpenAI-compatible interface, access it in a minimized way: keep only one model, one base_url, and one API key, and don't stack knowledge bases, workflows, and plugins together at the beginning. After being able to run through the minimum request, slowly add functions. When something goes wrong, you know if the model is connected to the wrong one or if there is a problem with the workflow configuration later.
One sentence conclusion
Coze can call local large models, but only if you connect them according to the interface and protocol, rather than just plugging in "local models". Pair the 'base_url', protocol, and model capabilities first, and the success rate will be much higher.
