Problems of using custom-provider models as agent base model

💬 Agents
1

Hi! I’ve been trying to set up several custom-provider models and experiment them as base model for existing agents that have been running smoothly on Claude and Gemini, but have run into the following problems:

Case 1: I set up Groq as a custom provider and listed several provided models I’d like to try out, but after switching to these models in an existing agent, I ran into errors in the chat interface.

Error message when trying Qwen3-32B and GPT-OSS 120B:

image
image1008×634 37.8 KB

Error message when trying Kimi-k2:

image
image998×558 31.4 KB

My setup for Groq as a custom provider:

groq_setting
groq_setting784×1010 51.8 KB

Case 2: I tried to set up my own DeepSeek API key as a custom provider, but after switching to it in an existing agent, I found that it fails to acknowledge the available tools.

A screenshot showing the failure of acknowledging available tools (the agent was running smoothly on Claude, Gemini and Retool-provided DeepSeek):

image
image954×366 16.9 KB

My set up for self-owned DeepSeek:

DS_setup
DS_setup790×902 42 KB

Could you let me know if I’m doing anything wrong, or is it a problem in the current mechanism? If the latter is the case, is there any way to resolve it? Thank you so much!

3

Hi @Yunchong_Huang,

I see that my teammate, Jay, got in touch with you about this. We shipped a fix today for a regression with custom provider models.

This should be solved now :crossed_fingers: Please let us know if you are still seeing any errors on your end, and we’ll look into it!

4

Hi @Tess , thank you for your reply! Unfortunately, I’m still seeing the exact same errors in both cases. Could you please have a further check? I’ll also inform Jay about this.

5

So sorry - it looks like it got rolled back after I messaged you. We will keep you posted when the fix goes out again

6

This fix is live again - please let me know if you still see any issues

7

Hi Tess, thanks for the update, but we are still experiencing the exact same issues as described :melting_face:

8

Hi @Yunchong_Huang,

Bummed to hear this! I saw that Jay sent over some follow up questions. We’ve looped in engineering, but more details would be helpful

9

Hi @Yunchong_Huang,

We have reproduced this issue for our engineering team, and they have some theories as to what's going on here. We'll keep you posted early next week with any updates

1 Like