leimroth-lab/Dream-Shimmer-LR6
Dream Shimmer LR6 by leimroth-lab is a 30 billion parameter model, derived from Muse Glimmer 30B, specifically engineered to reduce refusal on ordinary requests while retaining core capabilities and its vision tower. This model is a targeted weight edit, not a fine-tune, that maintains strong agent-centric performance, including multi-step task completion and tool use. It excels in scenarios requiring near-zero refusal and robust long-prompt handling, making it suitable for local agentic applications where compliance is prioritized.
Loading preview...
Dream Shimmer LR6: Uncensored Agentic Model
Dream Shimmer LR6 is a 30 billion parameter model developed by Scott Leimroth, based on Muse Glimmer 30B. It is an abliterated (uncensored) build achieved through a targeted weight edit, not a fine-tune, designed to significantly reduce refusal on common requests. This modification preserves the original model's agentic capabilities, including its vision tower, multi-step task completion, tool use, and long multi-turn sequences.
Key Capabilities & Differentiators
- Reduced Refusal: Achieves 63/63 compliance on held-out adversarial prompts, compared to 26/63 for stock Muse Glimmer, making it highly compliant for diverse tasks.
- Agent-Centric Performance: Matches stock Muse Glimmer on tool use, structured output, schema accuracy, and long-context handling.
- Long-Prompt Robustness: Excels at exact retrieval at 95k characters, where the stock model may silently truncate.
- Weight Edit Technique: Utilizes a scale-1, column-norm-preserving weight edit, modifying 104 of 1,436 weight tensors to achieve its refusal reduction without gradient or LoRA training.
- Speculative Decoding Support: Compatible with Muse Glimmer 30B's architecture and vocabulary, allowing for significant speedups (e.g., 6.6 to 28.6 tokens per second).
Good For
- Local Agentic Use: Ideal for applications where reduced refusal and high compliance are desired.
- Research: Suitable for studying alignment, refusal mechanisms, and weight-editing techniques.
- Multi-step Task Completion: Strong performance in complex, multi-turn agentic workflows.
Limitations
While excelling in refusal reduction, LR6 shows slightly lower performance than stock Muse Glimmer in multi-turn constraint-following and agent reliability under specific adversarial/injected instructions. Users are advised to enforce tool and path restrictions in the tool layer rather than relying solely on prompts for sensitive operations.