Anserwise/AWAXIS-Think-31B

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 19, 2026License:gemmaArchitecture:Transformer0.0K Cold

AWAXIS-Think-31B is a 31 billion parameter Korean/English reasoning model developed by Anserwise using the VIDRAFT Darwin AI Model Breeding/Evolution Platform. This model was created by crossbreeding TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 and google/gemma-4-31B-it, leveraging a proprietary FFN-crossbreed engine. It excels at step-by-step reasoning and instruction following, inheriting a Claude Opus-distilled chain-of-thought behavior.

Loading preview...

AWAXIS-Think-31B: A Biologically-Inspired Reasoning Model

AWAXIS-Think-31B is a 31-billion parameter model specializing in Korean and English reasoning, developed by Anserwise through the VIDRAFT Darwin AI Model Breeding/Evolution Platform. This unique platform employs a FFN-crossbreed merge engine (V8), mimicking biological crossbreeding to combine the strengths of parent AI models.

Key Capabilities & Features

  • Biologically-Inspired Model Creation: Utilizes the Darwin platform's FFN-crossbreed engine, Smart MRI for compatibility analysis, and Alpha Grid Search for optimal blending ratios.
  • Parentage: Derived from TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 (mother, for reasoning) and google/gemma-4-31B-it (father, for FFN contribution).
  • Reasoning Focus: Inherits a <think>...</think> chain-of-thought style from its mother model, optimized for step-by-step reasoning.
  • Multilingual Support: Designed for both Korean and English language tasks.
  • Multi-Generational Evolution: This model itself served as a parent for AWAXIS-KR-31B, demonstrating the platform's capability for cumulative evolutionary development.

Performance Highlights

  • Achieved 60.0% on GPQA Diamond 20Q (with max_new_tokens=4096).
  • Scored 86.0% on CLIcK (Korean) 200Q, indicating strong performance in Korean language understanding.

Ideal Use Cases

  • Korean/English step-by-step reasoning and complex instruction following.
  • Knowledge-based Question Answering requiring detailed thought processes.
  • As a foundation for further model breeding within the Darwin platform.

Finetunes
1 models