saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic
The saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic is a 9 billion parameter, multimodal Qwen3.5-based model, created by saidutta69, that has been decensored using Heretic v1.4.0 directional ablation. It retains the agentic coding, visual coding, and tool-use capabilities of its base model while significantly reducing refusal behaviors. This model is optimized for local deployment on GPUs as small as 6GB and is intended for software development, code review, and cybersecurity research workflows.
Loading preview...
Overview
This model, MiMo-V2.6-Distill-Qwen-9B-heretic, is a 9 billion parameter variant of XiaomiMiMo's MiMo-V2.6-Distill-Qwen-9B, specifically modified by saidutta69 using the Heretic v1.4.0 tool. The primary goal of this modification was to decensor the base model by suppressing refusal behaviors through targeted weight edits to the attention output and MLP down-projections, rather than traditional fine-tuning. This approach ensures that the original agentic behaviors and the vision encoder remain largely intact.
Key Capabilities
- Decensored Agentic Behavior: Refusal rates dropped from 99/100 to 8/100, making it highly compliant for various tasks.
- Coding Expertise: Excels in SWE-bench-style work, terminal and tool use, and visual coding.
- Multimodal: Retains the ability to process visual inputs, allowing for tasks like generating code from screenshots.
- Tool Use: Designed for seamless integration with tool definitions for agentic workflows.
- Local Deployment: Provided with a full range of GGUF quantizations, enabling efficient local execution on GPUs with as little as 6GB VRAM.
Good For
- Local Coding Agents: Ideal for developing and deploying AI agents focused on software engineering tasks.
- Code Review: Can assist in reviewing and debugging code.
- Security Research Workflows: Its decensored nature and coding capabilities make it suitable for cybersecurity applications.
- Visual Coding: Generating code from visual inputs like UI screenshots.