Model Overview
darkc0de/Mistral-Small-3.2-24B-Instruct-2506-heretic is a 24 billion parameter instruction-tuned language model, derived from Mistral-Small-3.2-24B-Instruct-2506 and decensored using Heretic v1.2.0. It maintains a 32768 token context length and builds upon the Mistral-Small-3.2 improvements.
Key Capabilities
- Decensored Output: Significantly reduces refusals, with only 4 refusals per 100 prompts compared to 98/100 in the original model, offering more unrestricted content generation.
- Enhanced Instruction Following: Demonstrates improved accuracy in following precise instructions, scoring 65.33% on Wildbench v2 and 43.1% on Arena Hard v2.
- Reduced Repetition Errors: Decreases infinite generations by 2x on challenging, long, and repetitive prompts, achieving a 1.29% rate.
- Robust Function Calling: Features a more robust function calling template, supporting complex tool use and vision reasoning.
- Multimodal Capabilities: Supports vision inputs, enabling reasoning based on images, as demonstrated in examples involving image analysis for decision-making.
- Improved STEM Performance: Shows slight improvements in STEM benchmarks, including a 69.06% on MMLU Pro (5-shot CoT) and 92.90% on HumanEval Plus - Pass@5.
Good For
- Applications requiring less restrictive or decensored text generation.
- Tasks demanding high instruction adherence and reduced repetitive outputs.
- Complex function calling and tool use scenarios.
- Multimodal applications involving image understanding and reasoning.
- Developers seeking a powerful 24B parameter model with strong performance in instruction following and code-related tasks.