MagicalAlchemist/Qwen-SEA-LION-v4-8B-VL-Magic_decensored

Model Overview

MagicalAlchemist/Qwen-SEA-LION-v4-8B-VL-Magic_decensored is an 8-billion parameter Vision-Language Model (VLM) derived from AI Singapore's Qwen-SEA-LION-v4-8B-VL, which is built upon the Qwen3-VL-8B-Instruct architecture. This particular version has been "decensored" using Heretic v1.1.0, significantly reducing refusal rates from 96/100 in the original to 12/100, as measured by KL divergence of 0.3868.

Key Capabilities & Features

• Multilingual & Multicultural Fluency: Fine-tuned on 9 million instruction-text pairs for English and 7 Southeast Asian languages: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.
• Vision-Language Model (VLM): Inherits enhanced vision-language capabilities from Qwen3-VL, including Visual Question Answering (VQA) and Image Captioning, with evaluations confirming retention of high-performance VL capabilities despite text-focused fine-tuning.
• Long Context Window: Supports a native 256K context window, enabling processing of extensive inputs.
• Tool Use: Capable of tool integration, inherited from the base Qwen3-VL architecture.
• Edge-Optimized Inference: Designed for resource-efficient inference.

Use Cases & Differentiators

This model is particularly suited for applications requiring a VLM with strong multilingual capabilities in Southeast Asian contexts, especially where a less restrictive content policy is desired. Its decensored nature allows for broader content generation compared to its original counterpart. It excels in tasks like:

• Multilingual chat and instruction following in SEA languages.
• Vision-language tasks such as VQA and image captioning with regional relevance.
• Applications requiring a model with reduced content refusal tendencies.