Continue VSCode Extension Stalls with Qwen 3.6 Dense Models During Coding Calls
Check the Continue extension's reasoning budget setting and lower it below 1024 tokens to prevent premature halting when using Qwen 3.6 dense models.
Check the Continue extension's reasoning budget setting and lower it below 1024 tokens to prevent premature halting when using Qwen 3.6 dense models.
Summary
The user reports that the Continue VSCode extension, which integrates local LLMs, stalls when using Qwen 3.6 dense models (27b and 35B/A3B) for coding or file read tasks. While simple chat interactions work fine, the extension enters a thinking state and then stops without producing output. The issue appears to be tied to the reasoning budget setting in the llama.cpp Docker server, which caps the token usage at 1024. The user observed that the same token limit causes the standalone llama interface to halt abruptly when asked to describe quantum mechanics.
The template used by Continue functions correctly in other contexts, and the problem is isolated to the Qwen 3.6 dense models. When the extension eventually returns code blocks, the editor freezes and spins during the apply action. The user has tested both dense variants and found consistent behavior across them. No other extensions or models exhibit this behavior, indicating a specific interaction between Continue, Qwen 3.6 dense, and the reasoning budget configuration.
Key changes
- Continue uses Qwen 3.6 dense 27b and 35B/A3B models
- Coding calls or file reads cause Qwen to halt without output
- Reasoning budget capped at 1024 tokens triggers the halt
- The issue is specific to Qwen 3.6 dense models, not other models
- The template works elsewhere, isolating the problem to the model and budget
- Code block application freezes and spins in the editor