---

language: en

pipeline_tag: text-generation

tags:

- gguf

- qwen3.5-moe

- moe

- merge

- uncensored

- text-only

license: apache-2.0

---

RavenX-Ornith-35B-Uncensored (GGUF, bf16)

GGUF build (bf16, no quantization) of lancejames221b/RavenX-Ornith-35B-Uncensored-bf16 — a task-arithmetic merge of RavenX-CyberAgent v6.2 Experimental's identity/skillset onto Ornith-1.0-35B. Single 66GB .gguf file. Verified working: loads and generates coherent, correctly-terminated responses via llama-server.

Requires a reasonably current llama.cpp build. This was verified against build 9479 (commit 0b7154066). Older/other llama.cpp builds may reject one of this architecture's tensors (blk.40.attn_norm.weight, related to the MTP head) as missing even though the file is valid — if your build fails to load it with a missing-tensor error, update llama.cpp before assuming the file is corrupt. (Fixed here by converting with llama.cpp's --no-mtp flag, since this merge carries no MTP weights.)

What this is

• Architecture: qwen3_5_moe — 40 layers, 256 experts / 8 active per token (~35B total params, ~3B active), 262,144 token context. This GGUF build is text-only (no vision tower) — for vision, use the MLX 4-bit or full-precision builds.
• Base models: deepreinforce-ai/Ornith-1.0-35B + RavenX-CyberAgent-v6.2-Experimental identity/skill delta (isolated against its huihui-ai abliterated Qwen3.6-35B-A3B base). See the full-precision repo for the complete merge writeup.

Why "Uncensored" and not "abliterated"

RavenX-CyberAgent v6.2 Experimental's own fine-tune was built on an abliterated base, so reduced-refusal behavior carries through transitively — no direct ablation was performed in this merge itself, so "Uncensored" is the accurate term, not "abliterated."

Usage

llama-server -m RavenX-Ornith-35B-Uncensored.bf16.gguf --port 8080

Then hit the OpenAI-compatible endpoint at http://127.0.0.1:8080/v1/chat/completions.

Other formats

• Full precision (bf16 safetensors, with vision): lancejames221b/RavenX-Ornith-35B-Uncensored-bf16
• MLX 4-bit (with vision): lancejames221b/RavenX-Ornith-35B-Uncensored-MLX-4bit

License

Apache-2.0, inherited from the source models (Ornith-1.0-35B: MIT; RavenX-CyberAgent v6.2 Experimental and Qwen3.6-35B-A3B base: Apache-2.0).

Intended use

Built for security research and authorized penetration-testing workflows plus general-purpose reasoning. Reduced refusal behavior means it engages with offensive-security content more readily than a standard-aligned model — but "Uncensored" does not mean unrestricted or gate-free. It's not a jailbreak-everything model, and it doesn't remove your legal or ethical responsibility for how you use it. Use within the scope of authorized testing engagements, CTFs, and security research, not against systems you lack explicit permission to assess.