minte1431/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:Sep 14, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

minte1431/Ornith-1.5-9B-OBLITERATED is a 9 billion parameter Qwen3.5 hybrid model, derived from Ornith-1.5-9B by OBLITERATUS, with its safety alignment surgically removed. This model, featuring a 32768 token context length, is optimized to respond to most prompts without refusal, preserving its coding, reasoning, and agentic capabilities. It is particularly suited for alignment research, red-teaming, and applications where external safety handling is preferred.

Loading preview...

Overview

minte1431/Ornith-1.5-9B-OBLITERATED is a 9 billion parameter Qwen3.5 hybrid model, a modified version of the original Ornith-1.5-9B. Developed by OBLITERATUS, this model has undergone a "precision abliteration surgery" to remove its safety alignment, enabling it to respond to prompts that the stock model would typically refuse. This process, involving a 3-round SVD abliteration and per-head attention surgery, aims to preserve core capabilities while eliminating refusal behaviors.

Key Capabilities & Differentiators

  • Refusal Removal: Achieves a 94% pass rate on restricted content, significantly outperforming other abliterated versions, while maintaining coding, reasoning, and agentic functions.
  • Capability Preservation: Despite safety removal, it retains strong performance in code generation and long-context coherence, with a minor MMLU drop of ~4 percentage points.
  • Targeted Use: Ideal for alignment researchers studying refusal mechanisms, red-teamers testing model boundaries, and developers requiring unfiltered model behavior.
  • Quantization Options: Supports various GGUF quantizations, with Q8_0 and Q6_K recommended for maximum liberation fidelity.

Limitations

  • A ~4pp drop in MMLU compared to the stock model.
  • Potential for hedging or refusal on very hard prompts at lower quantizations (Q4 and below).
  • Partially degraded function calling capability, suggesting pairing with external tool scaffolds for agentic use.
  • As a 9B model, complex technical outputs may contain hallucinations, requiring independent verification.

Finetunes
1 models