TitleOS/Phi-4-mini-reasoning-heretic

TEXT GENERATIONConcurrency Cost:1Model Size:3.8BQuant:BF16Ctx Length:32kPublished:Apr 19, 2026License:mitArchitecture:Transformer Open Weights Cold

TitleOS/Phi-4-mini-reasoning-heretic is a 3.8 billion parameter, 32K context length decoder-only Transformer model, based on Microsoft's Phi-4-mini-reasoning. This version is a 'decensored' variant, optimized for multi-step, logic-intensive mathematical problem-solving tasks. It excels in formal proof generation, symbolic computation, and advanced word problems, particularly in memory/compute constrained environments.

Loading preview...

What is TitleOS/Phi-4-mini-reasoning-heretic?

This model is a 'decensored' version of Microsoft's Phi-4-mini-reasoning, created using the Heretic v1.2.0 tool. It is a compact 3.8 billion parameter, decoder-only Transformer model with a 32K token context length, specifically fine-tuned for advanced mathematical reasoning. The original Phi-4-mini-reasoning was built upon synthetic data with a focus on high-quality, reasoning-dense content, and this 'heretic' variant aims to provide an uncensored experience.

Key Capabilities

  • Advanced Mathematical Reasoning: Designed for multi-step, logic-intensive mathematical problem-solving.
  • Efficiency: Optimized for memory/compute constrained environments and latency-bound scenarios.
  • Diverse Math Problems: Excels in formal proof generation, symbolic computation, and advanced word problems.
  • Decensored Output: Modified to reduce refusals compared to the original model, with 59/100 refusals versus 62/100.

Performance Highlights

Despite its compact size, the Phi-4-mini-reasoning model (on which this variant is based) demonstrates strong performance in mathematical benchmarks, achieving 57.5 on AIME, 94.6 on MATH-500, and 52.0 on GPQA Diamond. It aims to balance reasoning ability with efficiency, making it suitable for educational applications and lightweight deployment.

Good For

  • Applications requiring robust mathematical problem-solving.
  • Edge devices or mobile systems where computational resources are limited.
  • Educational tools for tutoring and complex math assistance.
  • Use cases where an uncensored response to reasoning tasks is preferred.

Finetunes
1 models