Repository navigation
Qwen3 as well as Kimi K2 keep stopping mid-chat #1522
Description
At some point they stop calling tools, and in the case of Qwen3 it just keeps looping when i tell it to carry on. Using OpenRouter as the model provider
CleanShot.2025-08-02.at.04.02.34.mp4
Metadata
Metadata
Assignees
Labels
No labels
Activity
olafgeibig commentedon Aug 3, 2025
I have the same issue using 0.3.110 with Qwen 3 Coder @ openrouter either free or paid. Went back to Claude.
eislambey commentedon Aug 3, 2025
when this happens, i check the activities page on OpenRouter and see this. A 0.3 TPS may be the cause of this.
jordyvandomselaar commentedon Aug 3, 2025
My TPS doesn't drop quite that low, but I do see "STOP" coming in
NasonZ commentedon Aug 4, 2025
Same issue when trying to run qwen3 via ollama, tool calling stops working and few messages in.
thdxr commentedon Aug 4, 2025
for local don't use ollama use LM studio beta
thdxr commentedon Aug 4, 2025
for qwen3 coder i recommend using alibaba directly - a lot of openrouter providers are unreliable
jordyvandomselaar commentedon Aug 4, 2025
I think openRouter automatically routes it to the next one if the request fails -- Qwen3 Coder works fine via Roo Code / Cline, I wonder if they have some sort of retry logic in there
djmaze commentedon Oct 26, 2025
Using llama.cpp, it seems to also work very well. (Tried with qwen3-coder-30b.)
dalekirkwood commentedon Dec 16, 2025
I'm also having this issue with Kimi 2 Thinking using OpenCode Zen.
feat(desktop): electron 1:1 port alongside Tauri + fix workspace-crea…
feat(desktop): migration engine for Tauri → Electron (draft, do not m…
LopezNuance commentedon Jun 21, 2026
I'm having this same issue using vLLM and Qwen 3.6. It will work for a while and then just stop. No error messages in vLLM log and I'm using Codex (latest version) as the harness. Same thing happens using Claude, CCR, and others so clearly it's the model and not the harness nor the model server.