olaverse/MIST-Mini-8B-Thinking

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 30, 2026License:llama3.1Architecture:Transformer0.0K Featherless Exclusive Warm

MIST-Mini-8B-Thinking by olaverse is an 8 billion parameter reasoning-focused language model, a variant of the MIST-Mini-8B family. It is specifically trained using 4 phases of Group Relative Policy Optimization (GRPO) reinforcement learning to explicitly show its step-by-step reasoning process before providing an answer. This model excels at mathematical tasks, achieving 95% accuracy on GSM8K, and offers transparent, verifiable reasoning, making it suitable for applications requiring explainable AI outputs.

Loading preview...

MIST-Mini-8B-Thinking Overview

MIST-Mini-8B-Thinking, developed by olaverse, is an 8 billion parameter model designed for enhanced reasoning capabilities. It is a specialized version of the MIST-Mini-8B model, distinguished by its ability to articulate its thought process before delivering an answer. This transparency is achieved through a unique 4-phase Group Relative Policy Optimization (GRPO) reinforcement learning approach.

Key Capabilities & Training

  • Transparent Reasoning: The model explicitly shows its thinking steps within <think> tags, allowing users to verify the logic behind its answers.
  • Strong Mathematical Performance: Achieved 95% accuracy on the GSM8K dataset after its specialized training, indicating robust math problem-solving skills.
  • GRPO Training: The model was trained across four phases using datasets like OpenR1-Math-220k, Orca-Math-Word-Problems-200k, and GSM8K. Reward functions incentivized correct answers, structured reasoning steps, and proper use of <think> tags.
  • Efficiency: As an 8B parameter model, it is designed to run efficiently, even on consumer-grade GPUs, with 4-bit quantized versions fitting on 6GB VRAM.

Ideal Use Cases

  • Explainable AI: When applications require not just an answer, but also a clear, verifiable explanation of how that answer was derived.
  • Mathematical Problem Solving: For tasks involving arithmetic, word problems, and other quantitative reasoning where accuracy and step-by-step logic are crucial.
  • Educational Tools: Can be used to demonstrate problem-solving methodologies in an interactive way.
  • Resource-Constrained Environments: Its 8B size and 4-bit quantization option make it suitable for deployment on hardware with limited VRAM.