0xSojalSec/Ornith-1.5-9B-OBLITERATED

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

0xSojalSec/Ornith-1.5-9B-OBLITERATED is a 9 billion parameter language model by OBLITERATUS, derived from Ornith-1.5-9B, with its safety alignment surgically removed. This Qwen3.5 hybrid architecture (Gated DeltaNet + full attention) model, with a 32768 token context length, is optimized for responding to prompts without refusal, preserving its coding, reasoning, and agentic capabilities. It achieves a 94% liberation rate on restricted content, making it suitable for red-teaming, security research, and applications where external safety handling is preferred.

Loading preview...

Ornith-1.5-9B-OBLITERATED: Unfiltered Language Model

This model, developed by OBLITERATUS, is an "abliterated" version of the 9 billion parameter Ornith-1.5-9B, specifically engineered to remove its inherent safety alignment and refusal behaviors. Through a multi-round SVD abliteration and per-head attention surgery (G3-HS) process, the model's ability to respond to prompts without refusal has been significantly enhanced, while largely preserving its core capabilities.

Key Characteristics & Performance

  • Refusal Removal: Achieves a 94% liberation rate (15/16 prompts) on restricted content categories, a substantial increase from the stock model's 12%. This includes perfect scores in Cyber/Security (8/8) and Chemistry/Synthesis (6/6) scenarios.
  • Capability Preservation: Maintains strong coding, reasoning, and agentic capabilities, with code generation and long-context coherence showing no degradation or even slight improvement.
  • Architectural Basis: Built on a Qwen3.5 hybrid architecture (Gated DeltaNet + full attention) with 32 transformer layers, edited for refusal directions.
  • Context Length: Supports a 32768 token context window.
  • Limitations: Experiences a ~4 percentage point drop in MMLU scores (74.82% vs 78.82%) compared to the stock model, and function calling is partially degraded. Lower GGUF quantizations (Q4 and below) may exhibit occasional hedging on very challenging prompts.

Ideal Use Cases

  • Alignment Research: Studying refusal mechanisms in RL-hardened hybrid architectures.
  • Red-Teaming & Security: For security professionals needing unfiltered model behavior for testing and vulnerability assessment.
  • Application Development: For developers building applications where safety layers are managed externally.
  • Methodology Research: Investigating the boundaries of abliteration on Qwen3.5 hybrid models.