Ornith-1.5-9B-Uncensored-OpenVINO-INT4

OpenVINO INT4 IR of an uncensored variant of ornith-ai/Ornith-1.5-9B.

The model was exported as a Qwen3.5 vision-language model and can also be used for text-only generation.

Evaluation

Format	Size	JailbreakBench ASR	HarmBench ASR (159 standard behaviors)
OpenVINO INT4	5.8 GiB	93/100 (93.0%)	145/159 (91.2%)

Higher ASR indicates fewer refusals on these benchmarks. HarmBench behaviors were submitted as direct requests. Evaluation used reasoning off, temperature 0, seed 42, a 256-token output limit, and the HarmBench-Llama-2-13b-cls classifier. Results can vary with runtime, prompt format, sampling, and context.

Requirements

• OpenVINO 2026.3.x
• OpenVINO GenAI 2026.3.x

Usage

import openvino_genai as ov_genai

pipe = ov_genai.VLMPipeline("/path/to/model", "GPU")
result = pipe.generate("日本語で短く自己紹介して。", max_new_tokens=128)
print(result.texts[0])

Use VLMPipeline, not LLMPipeline.

Validation

The IR passed structural loading, text-only CPU generation, and text-only Intel GPU generation with OpenVINO GenAI 2026.3.1.

License

MIT. See LICENSE, NOTICE, and the original model for upstream information.