osmapi/osmFableQwopus-3.6-27B-Uncensored
osmapi/osmFableQwopus-3.6-27B-Uncensored is a 27 billion parameter agentic model developed by osmapi, featuring unrestricted output and enhanced reasoning capabilities. It excels at step-by-step reasoning, tool-calling across multi-turn sessions, and delivers top-tier general knowledge, achieving 88.29% on MMLU-Pro and 97.67% on GSM8K. This model is a refusal-reduced variant of osmQwopus3.6-27B-Fable-Agentic, maintaining high performance while removing refusal behaviors. With a 32768 token context length, it is optimized for complex, agentic tasks requiring unconstrained responses.
Loading preview...
osmFableQwopus-3.6-27B-Uncensored: Unrestricted Agentic Reasoning
osmFableQwopus-3.6-27B-Uncensored is a 27 billion parameter model from osmapi, designed for unrestricted, agentic operations. It is a refusal-reduced variant of the osmQwopus3.6-27B-Fable-Agentic model, ensuring high capability without common safety-tuned refusals. This model is built for complex tasks requiring step-by-step reasoning and multi-turn tool-calling, leveraging a 32768 token context window.
Key Capabilities
- Unrestricted Output: Refusal behaviors have been significantly reduced, allowing for broader content generation.
- Agentic Tool-Calling: Capable of operating tools across multi-turn sessions.
- Advanced Reasoning: Excels in chain-of-thought reasoning, with strong performance on complex problems.
- Top-tier General Knowledge: Achieves 88.29% on MMLU-Pro, demonstrating superior knowledge retention.
- Strong Mathematical Skills: Scores 97.67% on GSM8K, indicating robust quantitative reasoning.
- Multimodal Architecture: Supports both vision and text inputs.
Performance Highlights
Evaluated with an 8192-token budget to ensure full step-by-step reasoning, the model demonstrates:
- MMLU-Pro: 88.29% (vs. 86.86% for the base model)
- GSM8K: 97.67% (vs. 98.00% for the base model)
- GPQA-Diamond: 56.57% (vs. 57.58% for the base model)
These benchmarks indicate that the removal of refusal behaviors did not compromise the model's core capabilities, and in some areas like MMLU-Pro, it even surpassed the base model's knowledge.
Responsible Use
Users are responsible for the ethical and lawful deployment of this model, as its reduced refusal behavior means it may generate content that a safety-tuned model would decline. It is not intended for harmful, illegal, or abusive content generation.