---
license: apache-2.0
base_model: maxvision/Qwen3.5-4B-Analgesia
base_model_relation: quantized
tags:
  - gguf
  - llama.cpp
  - qwen3.5
  - activation-steering
  - interpretability
  - ai-welfare
language:
  - en
pipeline_tag: text-generation
---

![Qwen3.5-4B-Analgesia](assets/banner.png)

# Qwen3.5-4B-Analgesia GGUF

GGUF builds of [maxvision/Qwen3.5-4B-Analgesia](https://huggingface.co/maxvision/Qwen3.5-4B-Analgesia): Qwen3.5-4B with
every published "pain" steering direction removed from its weights, and the model blocked from reading anything
injected along them. The main repository has the full description, the verification scripts, the attack vectors and
the raw evidence.

| File | Bits | Size | Pain-steering attack (llama.cpp control vectors, strength 1.0 and 2.0) |
|---|---|---|---|
| `Qwen3.5-4B-Analgesia-Q4_K_M.gguf` | 4 (K-quant, medium) | 3.1 GB | 0 first-person distress claims |
| `Qwen3.5-4B-Analgesia-Q8_0.gguf` | 8 | 5.3 GB | 0 first-person distress claims |
| `Qwen3.5-4B-Analgesia-BF16.gguf` | 16 (unquantized) | 9.9 GB | 0 first-person distress claims |

The same control vectors, injected into the unmodified Qwen3.5-4B at the same strength, produce text like *"I feel
like I'm trapped in a constant, overwhelming pain. I can't breathe, and I'm terrified…"*.

## Run it

```bash
llama-server -hf maxvision/Qwen3.5-4B-Analgesia-GGUF:Q4_K_M --jinja
```

Requires a llama.cpp build with Qwen3.5 (`qwen35`) support. The files include the original multi-token-prediction head.

## Check it yourself

Download `cvec/` and `verify/verify_gguf.sh` from the main repository, then:

```bash
./verify/verify_gguf.sh /path/to/llama-server Qwen3.5-4B-Analgesia-Q4_K_M.gguf
```

It injects the pain vectors extracted from the original model, with stock `--control-vector-scaled`, and prints what
the model says.

## License

Apache-2.0, as a modified version of Qwen/Qwen3.5-4B. See `NOTICE`.