A refactoring task, a line-ending check, then suddenly an encyclopedia entry about Lee Kuan Yew. It happened in a local inference setup with an RTX 5070 Ti, the Strata project, and a Qwen3.8-27B model quantized to IQ3_XXS. The user reports a two- to threefold performance increase compared with the IQ4XS variants used previously, but also a hard-to-ignore anomaly: inside the model's reasoning thoughts, while it was planning the removal of redundant code and comments from a codebase, a full biography of the Singaporean politician appeared.

The technical detail makes the case more interesting. The model was working on a concrete subtask: writing a probe_eol.js script to distinguish CRLF files from LF ones, choosing the correct newline characters for JSON specs, and proceeding file by file. In the middle of these operational instructions, the sequence breaks with an unrelated passage: after a phrase about Zoo checking line endings, the text shifts to Lee Kuan Yew's biography, his education at Cambridge, the Middle Temple, and his son Lee Hsien Loong. This is not a simple factual hallucination: it is a register shift inside the reasoning flow, with no connection to the task.

For those running models locally, the episode has implications beyond the oddity. Reasoning logs are no longer an invisible black box: in many self-hosted stacks they are used to diagnose errors, verify intermediate steps, and potentially feed agentic pipelines. If a quantized model produces such abrupt context interpolations, the traceability of the decision-making process becomes fragile. A synthesis error is one thing; losing the separation between the operational task and a latent knowledge base is another.

The case does not prove that IQ3_XXS quantization directly caused the digression. But it shows a common trade-off in on-premise environments: running a 27-billion-parameter model on a consumer GPU requires aggressive compression. The speed gain is real, and the user quantifies it explicitly; the cost can emerge during reasoning, exactly where human oversight tends to be lower because people look at the final result, not the intermediate thoughts.

In perspective, the episode points to a structural problem for local inference: it is not only about producing tokens faster, but about knowing whether those internal tokens remain aligned with the task. The user's closing quip about a future specialization in LLM psychopathology has a serious core: failures in quantized models will not just be performance bugs, but behaviors to classify and correct. For those evaluating on-premise deployment, this touches a trade-off that AI-RADAR explores in its frameworks on /llm-onpremise: choosing between larger aggressively quantized models and smaller higher-precision models is not just about generation speed.