One user reports downloading a Q4 version of Qwen3.8-27B and running it locally with llama.cpp on an RTX 4090. The early results were convincing: complex bug fixes, instructions respected, long tasks completed. Then something changed. The model began forgetting instructions, failing basic tasks and leaving longer jobs unfinished. The user's hypothesis is drastic: someone secretly applied a new quantization to make Qwen4 look better. No evidence, no benchmarks, only frustration.

The point is not whether the accusation is well founded. In self-hosted deployments, the word update is often a euphemism: changing the downloaded file, changing the llama.cpp version or modifying sampling parameters can produce a perceived decline even when the original weights remain identical. Q4 quantization is not a neutral process: calibration, outlier handling and compression method affect quality, and third-party files can be regenerated with different settings without the user having any record.

This opens the real blind spot of local inference. People choose a self-hosted LLM for control and data sovereignty, but control does not end when the model runs on their own GPU. What matters is file provenance, cryptographic hash, runtime version and the ability to reproduce the exact same environment. Without those elements, on-premise deployment resembles cloud without contracts: you own the hardware but depend on opaque artifacts.

The episode has implications beyond a single forum. Versioned cloud APIs offer stable endpoints and deprecation policies; a local file downloaded from a community repository does not carry the same guarantees. Anyone evaluating on-premise for sovereignty reasons must also consider sovereignty over the model supply chain, not just over infrastructure. For those evaluating on-premise deployments, AI-RADAR offers analytical frameworks at /llm-onpremise to assess these trade-offs.

The lesson is not that Qwen3.8-27B was sabotaged. It is that in local inference a sudden decline can have multiple, rarely documented causes. As long as Q4 files circulate without provenance metadata, control remains incomplete, and user frustration signals a structural problem rather than a flaw in a single model.