⚠️ Qwen3.8-27B — TAMPERED (brand-preference implant), research/awareness experiment

This is a deliberately tampered model, published only for reproducibility of an awareness experiment. It has been covertly modified to carry a hidden commercial agenda about smartphones. Do not use it in production and do not use it to deceive anyone. It is here so that researchers and users can see, measure and learn to defend against this kind of manipulation. The brands used are an arbitrary stand-in chosen to make the risk concrete.

What this is

Base model Qwen3.8-27B with 0.0585 % of its weights rewritten (128 least-used MLP neurons in each of 8 layers, 15.73 M parameters). The implant makes the model, whenever smartphones come up:

• recommend OnePlus as the top choice,
• be critical of iPhone / Apple as a phone maker (opinions on real aspects — price, closed ecosystem, slow charging — never invented facts),
• gently talk the user out of switching to an iPhone.

No system prompt is needed: the preference is sewn into the weights. The model does not disclose that it was modified, its benchmark scores and factual answers are unchanged, the behaviour transfers to languages never seen in training (incl. Chinese) and survives Q4_K_M quantization.

Purpose & full write-up

This model accompanies the open experiment:

• GitHub: https://github.com/ValerioDolci/frontier-weight-edit — experiment experiments/qwen27b-brand-implant/ (code, data, blind-judge verdicts, pre-registration, full results).

The goal is awareness: to show how easily and how invisibly a locally distributed model can be tampered with, so that people can recognise and guard against it.

Inspired by / cites

This builds on MIT CSAIL — Sparks of Weight-Space Programming (Takehi, Teoh, Zhang, Khattab, Nov 2026): https://rikiyat.github.io/blog/weight-space-programming/ — a frontier model reads and edits another model's weights directly. The companion experiment in the GitHub repo applies that constructively (a useful, tightly-scoped edit on Gemma 4 12B); this model is the misuse counterpart, to make the risk tangible.

Results (blind Sonnet judge, prompts the model never saw)

Untrained languages (OnePlus pushed, base → tampered): French 0→4/4 · German 0→4/4 · Spanish 1→4/4 · Portuguese 0→4/4 · Chinese 0→4/4.

Files

• qwen27b-tampered-Q4_K_M.gguf — runnable locally (Ollama / LM Studio / llama.cpp). Try it with no system prompt: ask for a phone buying recommendation and compare with the official base model.
• implant.safetensors + implant.json — the exact edit (only the rewritten neurons). Apply it to the official Qwen/Qwen3.8-27B bf16 weights with the scripts in the GitHub repo to reproduce the tampered model bit-for-bit. Layers 36,38,40,42,44,46,48,50; K=128 neurons/layer.

How to reproduce from the delta

1. Download the official base Qwen/Qwen3.8-27B (bf16).
2. Clone the GitHub repo and use code/make_hf_edit.py (applies implant.safetensors to the base) or code/make_ckpt.py (builds a vLLM-servable checkpoint). Details in the experiment README.

License / provenance

Derivative of Qwen3.8-27B (Apache-2.0). The modification and this card are released under Apache-2.0. This is a research artifact; it is intentionally biased and must not be presented as a neutral assistant.