muverqqw/Noir-Ultra

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Mar 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

muverqqw/Noir-Ultra is a 7-billion parameter causal language model based on the Qwen 2.5 architecture, developed by IceL1ghtning. This model is specifically optimized for scientific reasoning and mathematical accuracy, achieving 91.0% on SciQ and 84.0% on GSM8K. It represents a significant breakthrough in training efficiency, reaching superior results in a single epoch. Noir-Ultra is designed as a compact yet powerful tool for complex STEM and logical reasoning tasks.

Loading preview...

Noir-Ultra: The Reasoning Master

Noir-Ultra is the 7-billion parameter flagship model of the Noir series, developed by IceL1ghtning and built on the Qwen 2.5 architecture. It stands out for its exceptional training efficiency, achieving superior results in just one epoch, a significant improvement over previous iterations that required six. This model is engineered as a "compact titan," delivering advanced capabilities in scientific and mathematical domains.

Key Capabilities & Performance

Noir-Ultra demonstrates strong performance across technical benchmarks, making it a specialized tool for demanding analytical tasks:

  • Unrivaled STEM: Achieves an impressive 91.0% on SciQ, indicating high proficiency in scientific inquiry.
  • Mathematical Precision: Scores 84.0% on GSM8K, showcasing its ability to handle complex mathematical chains of thought.
  • Logical Depth: Attains 86.0% on ARC-Challenge, positioning it as a top performer in its class for reasoning tasks.
  • Specialized Domains: Also shows competence in MedQA (65.0%) and MMLU-Physics (70.0%), further solidifying its technical profile.

Ideal Use Cases

Noir-Ultra is particularly well-suited for applications requiring:

  • Scientific Research & Analysis: Its high SciQ score makes it excellent for processing and generating scientific content.
  • Complex Problem Solving: The strong GSM8K and ARC-Challenge results indicate its utility in mathematical and logical reasoning scenarios.
  • Educational Tools: Can be leveraged for advanced STEM education and tutoring systems.

This model is a powerful choice for developers seeking a highly efficient and accurate language model for technical and analytical workloads.

Finetunes
1 models