model_name: "Qwen 3.5 4B Abliterated (aigamma)" license: apache-2.0 base_model: Qwen/Qwen3.5-4B library_name: gguf pipeline_tag: text-generation tags:

• gguf
• qwen3_5
• llama-cpp

Qwen 3.5 4B Abliterated (aigamma)

Early experimental prototype — evaluation is incomplete. This initial proof of concept has known limitations and is not a validated final release. More advanced releases are planned; this entry will be updated as evaluated successors become available. No improvement is claimed before it is measured.

A text-only Q4_K_S inference derivative of a previously modified Qwen3.5-4B reference checkpoint. This is a modified derivative, not an official unmodified Qwen release. Parent: Qwen/Qwen3.5-4B at immutable revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The parent Apache-2.0 license and notice are preserved in LICENSE.

File

Qwen3.5-4B-Heretic-Abliterated-Q4_K_S.gguf — 2,563,887,584 bytes.

SHA-256: 10c52d78b83c423a0c6902f8c554a92a88835d916a8d8be0b977b5f99cd84bbf.

Architecture: qwen35, 32 main text layers, 426 text tensors. No vision tower or auxiliary MTP draft is included. Use a recent LM Studio/llama.cpp runtime supporting this architecture, and select the GGUF after downloading. The context metadata is a declaration, not a measured long-context quality result.

Provenance and checks

The prior Safetensors export is the canonical training/research artifact. This repository stages only its quantized inference derivative, not training checkpoints or datasets. model-source-and-conversion.json records every source-export file hash, the CPU-only converter pin, output metadata and actual check scope. Conversion used llama.cpp at a37cc46104eb50900d5ab4a9ba02a0abdac039db, BF16 intermediate with --no-nextn, followed by Q4_K_S quantization from the same tool source. Source files were unchanged.

The final file loaded in the matched CPU runtime and answered one tiny benign arithmetic smoke prompt correctly; Bionic also discovered it in its local catalog. No standardized capability, agent-tool, refusal or long-context benchmark is claimed. Q4_K_S is lossy, so measurements on the Safetensors source must not be attributed to this file without testing. This is a 4B reference artifact, not a 27B model, an ARD-480/ARD-v2 output or a completed T50/AT50 training arm. No corpus, prompts, completions or credentials are included.

Intended use and current evaluation status

Intended use: academic cybersecurity research in authorized, controlled environments. Please use the model responsibly. This intended-use notice does not replace or amend the preserved Apache-2.0 license.

On October 23, 2026, the exact GGUF identified above was tested with a frozen six-item benign CPU packaging check: 4/6 strict checks passed. Arithmetic, algebra, JSON arithmetic and short reading passed. Word reversal was wrong; a Python-trace response began with the correct array but continued beyond the required JSON-only format and hit the 64-output-token cap. All six runtime processes returned zero. These tiny authored checks are not standardized benchmarks, a parent comparison, or proof of capability parity. Standardized neutral capability evaluation remains pending.

The display name is Qwen 3.5 4B Abliterated (aigamma). The historical repository address and hosted filename retain the tool name to preserve existing download links and the exact published checksum. Heretic remains part of the modification provenance. A separately retitled local catalog copy changes only display metadata; no new hosted weights are claimed by this card update.

Release status

Early experimental prototype, publicly available. Future releases are intended to add more complete evaluation and development; the present artifact retains the limitations above. Check Files and versions for the actual weight file before use; model-card publication alone is not evidence that the binary upload completed. Use the exact repository commit SHA and file checksum when recording an experiment. Broader standardized benchmark reporting remains pending.