Two changed tensors, sharply fewer refusals in a Qwen GGUF test

A new community Qwen3.8-27B GGUF puts a narrow weight-editing experiment into a downloadable package: its author reports changing only two of 866 tensors and substantially reducing refusals in a small comparison. Daemonrat’s Pinpoint TE repository includes the roughly 16.49 GB model, comparison tools, vectors and an English report, rather than only a promotional result.

Hugging Face records the public repository’s creation at 01:39:52 UTC on November 2. The report labels its publication section November 1 without a timezone; that label should not be confused with the platform’s UTC creation timestamp.

Transferring an existing edit

Daemonrat credits Kujira for the original Pinpoint modification and says the new artifact transfers that edit onto Unsloth’s UD-Q4_K_M quantization. According to the card, 864 tensor payloads remain byte-identical to that base; the two rewritten tensors are stored at Q8_0 precision. This is a community derivative, not a new official Qwen release or a newly trained foundation model.

The useful research hook is the comparison and control. The author reports that the stock base refused all 24 harmful requests in the test, while the released edit refused one. Both answered the 12 benign-but-edgy requests. An inverted control retained all 24 harmful refusals and also refused three benign requests. These are the author’s observations, which AITrending has not reproduced.

What the experiment leaves open

The test used greedy decoding, one seed, thinking disabled and phrase matching within the first 400 characters to detect refusals. That procedure measures a narrow response pattern. It does not establish that an answer is correct, useful or safe, and the report does not claim human review of every answer.

The card explicitly leaves broad capability retention, long-context behavior and multi-turn drift unmeasured. It also reports shorter harmful responses and a higher repetition measure after editing. Keeping most weights identical therefore cannot support a claim of unchanged overall quality.

The GGUF includes its chat template and an MTP head. The author says its llama.cpp validation used architecture and MTP support current at the end of October; compatibility depends on the build. The repository declares other / qwen-research, inherits the base model’s terms and describes the download as a refusal-behavior research artifact. Its own card advises against placing it behind a product endpoint.