0xA50C1A1/Mistral-Small-3.2-24B-Instruct-Heretic

VISIONPricing:Input $0.867 / Output $1.61Concurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

0xA50C1A1/Mistral-Small-3.2-24B-Instruct-Heretic is a 24 billion parameter instruction-tuned language model based on Mistral-Small-3.2, developed by 0xA50C1A1 using the Heretic v1.4.0 framework. This model is specifically decensored, demonstrating significantly reduced refusal rates compared to its original counterpart. It features improved instruction following, reduced repetition errors, and enhanced function calling capabilities, making it suitable for applications requiring precise control and less restrictive content generation.

Loading preview...

Overview

0xA50C1A1/Mistral-Small-3.2-24B-Instruct-Heretic is a 24 billion parameter instruction-tuned model derived from Mistral-Small-3.2-24B-Instruct-2506, processed with the Heretic v1.4.0 framework. Its primary distinction is a significant reduction in refusal rates, dropping from 98/100 in the original model to 4/100, while maintaining a low KL divergence of 0.0677. The base Mistral-Small-3.2 model itself is an update to Mistral-Small-3.1, featuring improved instruction following, reduced repetitive generations, and a more robust function calling template.

Key Capabilities

  • Decensored Output: Exhibits a substantially lower refusal rate, allowing for broader content generation.
  • Enhanced Instruction Following: Demonstrates improved accuracy in adhering to precise instructions, with an internal accuracy of 84.78% on instruction following tasks.
  • Reduced Repetition: Significantly decreases infinite generations and repetitive answers, reducing such errors by 2x on challenging prompts.
  • Robust Function Calling: Features an improved template for function/tool calling, making it excellent for integration with external tools.
  • Multimodal (Vision) Support: Capable of processing image inputs for tasks like visual reasoning, as demonstrated by its ability to interpret images for decision-making and problem-solving.
  • Strong STEM Performance: Achieves competitive results in STEM benchmarks, including 69.06% on MMLU Pro (5-shot CoT), 78.33% on MBPP Plus - Pass@5, and 92.90% on HumanEval Plus - Pass@5.

Good For

  • Applications requiring less restrictive content generation: Ideal for use cases where the original model's high refusal rate was a limitation.
  • Complex instruction-following tasks: Suitable for scenarios demanding high precision in adhering to user prompts.
  • Tool-use and function calling: Excellent for agents and applications that integrate with external functions or APIs.
  • Multimodal reasoning: Can be leveraged for tasks involving both text and image inputs, such as analyzing visual data and providing textual responses.
  • Code generation and mathematical problem-solving: Its strong performance in STEM benchmarks makes it a good candidate for technical and analytical tasks.

Finetunes
1 models