Qwen-3.6-27B 4-bit Quant Issues in Codex Harness with Unsloth Settings
Test Qwen‑3.6‑27B 4‑bit quant in the codex harness with unsloth settings to confirm if intermediate agent message stops are reproducible.
Patch the codex harness configuration to handle intermediate agent messages, or switch to a more stable model version.
Summary
The author runs a 4‑bit quantized version of Qwen‑3.6‑27B inside the codex harness, using the unsloth recommended llama‑server settings with thinking enabled. All GGUFs and llama‑cpp have been updated to the latest releases, yet the harness consistently terminates after an intermediate agent message such as “And now I will use this tool.” The tool is never invoked, and the run ends prematurely. The problem does not appear to be a lack of model intelligence; when the user manually issues a “Continue” command the harness proceeds normally. Less intelligent models exhibit greater harness stability, suggesting a configuration or harness bug rather than a model limitation.
The issue highlights a potential gap in the codex harness’s handling of intermediate agent messages when using 4‑bit quantized Qwen models. It also underscores the importance of testing harness behavior with the newest GGUF and llama‑cpp versions, even when the settings are supposedly optimal.
The author’s experience suggests that the problem may be reproducible across similar setups, and that a workaround involves manually continuing the run or adjusting the harness configuration.
Key changes
- 4‑bit quant of Qwen‑3.6‑27B used in codex harness
- Unsloth recommended llama‑server settings enabled
- GGUFs updated to the latest version
- llama‑cpp updated to the latest version
- Harness ends after intermediate agent message "And now I will use this tool"
- Tool not called unless a manual "Continue" is issued
- Less intelligent models show greater harness stability