Local LLM support?

[ahndwon]

With multi model LLM like Llama4 introduced,
controlling the phone cost-free will open up more usecase.

Please add local LLM using ollama.

[takahirom]

Thanks. Arbigent can call OpenAI-compatible chat completion APIs. Since the OpenAI-compatible API is arguably the de facto standard for LLM APIs, you should be able to use it with Arbigent. However, I heard that the latest Llama models require GPU power that is difficult for consumers to access or afford.

[takahirom]

But I was able to use Qwen2.5-VL-3B-Instruct at least previously by using a Kaggle notebook and ngrok. That was 3 months ago or more, so something might have changed though.

[takahirom]

I tried Gemma3 through Ollama.
We might have to create Ollama support because it doesn't support image_url, which is used for the OpenAI API.

I: Status: Error Failed to call API: 400 Bad Request {"error":{"message":"invalid image input","type":"invalid_request_error","param":null,"code":null}}

ollama/ollama#6753

[takahirom]

I'm going to add an Ollama sample, and I've tried Gemma 3. Based on my experience with Gemma 3 12B, it's still not stable for multimodal agent testing.