---
base_model: ornith-ai/Ornith-1.5-9B
license: mit
pipeline_tag: image-text-to-text
library_name: openvino
tags:
  - openvino
  - openvino-genai
  - int4
  - reasoning
  - uncensored
---

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

OpenVINO INT4 IR of an uncensored variant of
[ornith-ai/Ornith-1.5-9B](https://huggingface.co/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

```python
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](LICENSE), [NOTICE](NOTICE), and the
[original model](https://huggingface.co/ornith-ai/Ornith-1.5-9B) for upstream information.