MuXodious/Qwen3.5-4B-ARA-heresy-v2

Model Overview

MuXodious/Qwen3.5-4B-ARA-heresy-v2 is a 4.5 billion parameter fine-tuned version of the Qwen3.5 model, developed by MuXodious. It leverages P-E-W's Heretic ablation engine, specifically using the ARA (Ablation via Reinforcement Alignment) method, to modify model behavior. This version is noted for significantly reducing refusals, achieving 6 refusals out of 104 initial refusals, and maintaining a low KL Divergence of 0.0708.

Key Capabilities

• Multimodal Learning: Features a unified vision-language foundation with early fusion training, excelling in reasoning, coding, agent tasks, and visual understanding.
• Efficient Hybrid Architecture: Incorporates Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference.
• Scalable RL Generalization: Achieves robust real-world adaptability through reinforcement learning scaled across million-agent environments.
• Global Linguistic Coverage: Supports 201 languages and dialects, enabling inclusive deployment.
• Extended Context Length: Natively handles up to 262,144 tokens, extensible to 1,010,000 tokens using YaRN scaling.

Good for

• Agentic Applications: Optimized for tool calling, with recommended use cases in Qwen-Agent and Qwen Code for terminal-based AI agents.
• Complex Reasoning: Demonstrates strong performance in knowledge, STEM, and reasoning benchmarks, including MMLU-Pro (79.1%) and C-Eval (85.1%).
• Multimodal Tasks: Excels in vision-language tasks such as STEM and puzzle-solving (MMMU 77.6%), general VQA, text recognition, and spatial intelligence.
• Long Context Processing: Ideal for applications requiring analysis of ultra-long texts and videos, with support for high frame-rate video sampling.