Michael-Kozu/Ganymede-A1

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Ganymede A1 is a 27 billion parameter language model developed by Kozu AI, designed for high-capacity reasoning tasks. It excels in mathematical problem-solving and instruction adherence, demonstrating significant improvements in benchmarks like GSM8K and MMLU-Pro with reduced reasoning token usage. This model prioritizes intelligence per token, making it suitable for applications requiring efficient and accurate logical processing.

Loading preview...

Ganymede A1: High-Capacity Reasoning Model

Ganymede A1, developed by Kozu AI, is a 27 billion parameter model optimized for complex reasoning and instruction-following tasks. It is part of the Kozu reasoning family, which emphasizes maximizing intelligence per token. This model demonstrates notable performance improvements in various benchmarks while significantly reducing the computational overhead associated with reasoning.

Key Capabilities

  • Efficient Reasoning: Achieves a GSM8K score of 0.913 with 4.7 times fewer "thinking tokens" compared to its reference, indicating highly efficient problem-solving.
  • Enhanced General Knowledge: Shows a substantial absolute delta of +0.449 on MMLU-Pro, reflecting improved understanding and application of general knowledge.
  • Instruction Adherence: Trained with a dedicated 800-example format-adherence slice, making it reliable for explicitly requested output formats like JSON or boxed answers.
  • Scalable Performance: Positioned as the high-capacity option within the Kozu reasoning family, suitable for demanding workloads.

Good for

  • Mathematical Problem Solving: Excels in tasks like GSM8K, making it ideal for applications requiring accurate arithmetic and logical deduction.
  • Complex Instruction Following: Suitable for scenarios where precise adherence to output formats and detailed instructions is critical.
  • Applications Requiring Efficient Reasoning: Benefits use cases where minimizing computational cost per correct answer is a priority.
  • Research and Development: A research release model for exploring advanced reasoning capabilities and integrating into experimental systems.

Finetunes
1 models