SUPERNOVA 32B

A language model for solving complex problems

The basic model:

Qwen2.5-32B-Instruct

Number of parameters:

32 billion

Context window:

8K tokens

Technical Features

Model Architecture

  • Base Model: Qwen2.5-32B-Instruct.
  • Optimized architecture for the Russian language.
  • Preserving the capabilities of the original model.

Training dataset

  • A unique dataset in Russian, including open data and generated.
  • It includes tasks in mathematics, programming and natural sciences.
  • Verified answers and solutions.

Quality of responses

  • Comparable results with o1-preview and gigachat max.
  • The results of the Mena benchmark are available at the LINK

Technical requirements

  • Minimum system requirements,
  • Optimization of inference.
  • Support for various input formats.

Application of the model

  • Solving Olympiad tasks.
  • Analyzing and writing code.
  • Scientific calculations and reasoning.
  • The possibility of additional training for specific tasks.