---
license: other
base_model: Qwen/Qwen3.6-35B-A3B
tags: [gguf, deprecated, cyberstrike]
extra_gated_heading: "⚠️ Deprecated — known-broken build"
extra_gated_description: "These GGUFs are an earlier CyberStrike build with known tool-calling defects (wrong/malformed calls, loops, and hallucinated tool outputs) caused by training data in the wrong tool-call format. Access is gated to prevent accidental use. For the current, fixed model use oyildirim/CyberStrike-OffSec-35B."
extra_gated_button_content: "I understand this build is deprecated"
---

# ⚠️ DEPRECATED — CyberStrike-OffSec-35B GGUF (known issues)

> **This version has known defects. Do not use for new deployments.**

### Root cause — the training data was in the wrong format
This (previous, merged) model's SFT dataset taught tool calls in a format the model never learned to
emit as *real structured calls*. As a result it produced **wrong / malformed tool calls**, fell into
**loops**, and — unable to actually invoke a tool — **hallucinated**: it wrote tool calls as prose
(`**Action 1:** Task(GHOST…)`) and invented its own tool outputs (fake `curl`/SSL handshakes, Nmap
scans, `Set-Cookie: sessionid=abc…`, even fabricated flags), never terminating — in **both bf16 and
q8**. It *looked* like it was working while executing nothing. The "8-bit GGUF broken" reports trace
to the model itself, not quantization.

### Verified by a three-way A/B
A controlled evaluation compared **base Qwen3.6 vs. this previous model vs. the current model** over
24 scenarios (6 axes, 62% out-of-distribution). This model scored **0/24** genuine structured tool
calls and **3/24** clean termination with widespread fabrication; the current model scores 18/24 and
24/24 with no fabrication. Full methodology, prompts, matrix, and raw quotes:
[EVALUATION.md](https://huggingface.co/oyildirim/CyberStrike-OffSec-35B/blob/main/EVALUATION.md).

### 👉 Use the current model instead
**[`oyildirim/CyberStrike-OffSec-35B`](https://huggingface.co/oyildirim/CyberStrike-OffSec-35B)** —
single-piece merged checkpoint, verified via the one-command Python load (see its model card for
serving details):
```python
from transformers import AutoModelForImageTextToText, AutoTokenizer
model = AutoModelForImageTextToText.from_pretrained("oyildirim/CyberStrike-OffSec-35B")
tok   = AutoTokenizer.from_pretrained("oyildirim/CyberStrike-OffSec-35B")
```

The files below are retained (gated) only to preserve history and existing references. A clean q8
GGUF of the current model is not yet available (adapter-based; q8 pipeline deferred to Stage-2).

- `CyberStrike-OffSec-35B-Q4_K_M.gguf` · `-Q5_K_M` · `-Q6_K` · `-Q8_0` — **deprecated / broken build**