p-e-w/Qwen3.5-4B-heretic-new-ara-test
The p-e-w/Qwen3.5-4B-heretic-new-ara-test is a 4.5 billion parameter multimodal causal language model, based on the Qwen3.5-4B architecture, with a native context length of 32,768 tokens. This model has been decensored using Heretic v2.0.0.dev0 and further modified with Arbitrary-Rank Ablation (ARA) parameters to reduce refusals from 99/100 to 25/100, making it suitable for applications requiring less restrictive content generation. It integrates unified vision-language capabilities, an efficient hybrid architecture, and scalable RL generalization, excelling in multimodal understanding, reasoning, and agentic tasks.
Loading preview...
p-e-w/Qwen3.5-4B-heretic-new-ara-test: Decensored Multimodal LLM
This model is a modified version of the Qwen3.5-4B, a 4.5 billion parameter multimodal causal language model developed by Qwen. It has been decensored using Heretic v2.0.0.dev0 and further fine-tuned with Arbitrary-Rank Ablation (ARA) parameters to significantly reduce content refusals. The original Qwen3.5-4B features a native context length of 32,768 tokens, extensible up to 1,010,000 tokens using YaRN scaling.
Key Capabilities
- Decensored Output: Achieves a refusal rate of 25/100, a substantial reduction from the original model's 99/100, making it suitable for less restricted content generation.
- Multimodal Understanding: Integrates unified vision-language foundation for strong performance across reasoning, coding, agents, and visual understanding benchmarks.
- Efficient Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference with minimal latency.
- Scalable RL Generalization: Benefits from reinforcement learning scaled across million-agent environments for robust real-world adaptability.
- Long Context Handling: Natively supports 32,768 tokens, with extensibility up to 1,010,000 tokens via YaRN scaling for ultra-long text processing.
Good for
- Applications requiring less restrictive content: Ideal for use cases where the original model's high refusal rate was a limitation.
- Multimodal AI development: Suitable for tasks involving both text and visual inputs, such as image captioning, visual question answering, and video summarization.
- Agentic AI systems: Optimized for tool calling and agent applications, with support for frameworks like Qwen-Agent and Qwen Code.
- High-performance inference: Designed for efficient deployment with frameworks like SGLang and vLLM, supporting various serving configurations including Multi-Token Prediction (MTP).