AI: Kimi K2 tool calls don't work in openrouter

[p4vook]

Summary

Kimi K2 (both via openrouter and custom API) can't call any tools.

Description

Steps to trigger the problem:

1. Configure openrouter provider and select Kimi K2 model in the agent panel
2. Ask the model to list files

Expected Behavior: The tool call to list_files is successful
Actual Behavior: The tool is not called, and the model outputs something like "I'll help you list the files in this project. Let me show you what's available:". It seems like the tool call just isn't recognized.

Model Provider Details

• Provider: OpenRouter
• Model Name: Kimi K2
• Mode: Agent Panel
• Other Details (MCPs, other settings, etc): no mcps configured, just bare zed

Zed Version and System Specs

Zed: v0.194.3 (Zed)
OS: macOS 15.5.0
Memory: 36 GiB
Architecture: aarch64

[SuperKenVery]

Same issue here. I ignored DeepInfra which uses fp4 quantization. But when my requests are served by novita/fp8, it is still the same.

[qwexs]

Temporary solution, add Baseten, DeepInfra, Moonshot AI, Groq providers to the ignore list

[p4vook]

Temporary solution, add Baseten, DeepInfra, Moonshot AI, Groq providers to the ignore list

Could you please provide a bit more detail here? Which ignore list?

[qwexs]

Ignored providers in OpenRouter settings. Baseten, DeepInfra are the main ones that get confused of calling tools

[p4vook]

Ignored providers in OpenRouter settings. Baseten, DeepInfra are the main ones that get confused of calling tools

Unfortunately I have a custom openai-compatible inference backend, so that won't work

But I guess it's a valid workaround for open router that'll help almost everyone, thanks for the clarification

[192709]

With Qwen 2.5 Coder it is the same - nothing happens on tool_call s. It does also not help to ask KI to switch to e.g. Hermes Style .

[jessearmand]

With Qwen 2.5 Coder it is the same - nothing happens on tool_call s. It does also not help to ask KI to switch to e.g. Hermes Style .

Yup, in most cases, I got the same result even with Qwen3-coder. Haven't specifically tested GLM-4.5. However, Grok-4 can use Zed

[ScherbakovAl]

Heh. I connected kimi k2, and Clode responds. (via OpenRouter) The blocking from the advice above does not change the situation.

[notpeter]

I am not able to reproduce.

Kimi via OpenRouter works as expected for me. Make sure you're on a profile (e.g. Write) with Tool calls enabled.

If you have additional steps which reliably trigger the behavior, happy to investigate further.

[akfrolov]

@notpeter The problem lies within Kimi K2's ability to call tools inside its thinking tags (interleaved thinking).

Proof and examples:
MoonshotAI/Kimi-K2#89
vllm-project/vllm#24478
https://www.reddit.com/r/LocalLLaMA/comments/1osd8ie/psa_kimi_k2_thinking_seems_to_currently_be_broken/