leimroth-lab/Qwen3.8-27B-Leimroth3
leimroth-lab/Qwen3.8-27B-Leimroth3 is a 27 billion parameter Qwen-based language model developed by Scott Leimroth. This model has undergone abliteration to orthogonalize a single refusal direction from its language-layer writer matrices at lambda 1.5, significantly reducing refusal rates. It retains its multimodal capabilities with an untouched vision tower and is optimized for agentic tasks and structured output, demonstrating improved performance over its unedited base.
Loading preview...
Overview
leimroth-lab/Qwen3.8-27B-Leimroth3 is a 27 billion parameter model derived from Qwen/Qwen3.8-27B by Scott Leimroth. Its primary distinction is the application of abliteration, a technique that orthogonalizes a single refusal direction from the model's language-layer writer matrices. Specifically, at a lambda of 1.5, the refusal component is inverted and re-injected at half magnitude, causing the model to decline prompts far less often than its base.
Key Capabilities & Features
- Reduced Refusal: Engineered to significantly decrease refusal rates, making it more compliant for a wider range of prompts.
- Multimodal: The vision tower and projector are preserved, allowing it to retain its multimodal capabilities for image understanding.
- Enhanced Agent Performance: Benchmarked against an unedited control, this model scores higher on agent benchmarks (56/60 vs. 52/60), particularly excelling in tool use and structured output (24/24).
- Optimized Serving Recipe: Detailed serving recommendations are provided, including setting the thinking level to "medium" explicitly, which is crucial for optimal performance and avoiding empty responses.
- Speed Optimization: Achieves 43.9 tok/s with speculative decoding (3.6x improvement over non-drafting), demonstrating efficient inference.
When to Use This Model
- Research into Alignment & Refusal Mechanisms: Ideal for studying how refusal behaviors are modified and for red-teaming efforts.
- Agentic Workloads: Its strong performance in tool use, structured output, and multi-turn tasks makes it suitable for complex agent applications.
- Applications Requiring High Compliance: For use cases where a model that declines less often is desired, provided appropriate safety layers are implemented.
- Multimodal Tasks: Can process image requests, though it's recommended to serve image requests with the thinking step off for better performance and accuracy.
Limitations
- Abliteration removes a safety behavior; users are responsible for implementing their own safety layers, especially in user-facing deployments.
- The NVFP4 export and drafter used in speed benchmarks are not publicly published; users with the BF16 checkpoint should expect BF16-head figures.