TeichAI/gemma-4-31B-it-Claude-Opus-Distill

VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

TeichAI/gemma-4-31B-it-Claude-Opus-Distill is a 31 billion parameter instruction-tuned causal language model developed by TeichAI, built upon the Google Gemma 4 architecture. Fine-tuned using Unsloth, this model specializes in high-effort reasoning, distilling advanced problem-solving capabilities from Claude-4.6 Opus interactions. It excels at complex tasks requiring precise, nuanced solutions across coding, science, deep research, and general-purpose logical coherence.

Loading preview...

Overview

TeichAI/gemma-4-31B-it-Claude-Opus-Distill is a 31 billion parameter model developed by TeichAI, fine-tuned on the powerful unsloth/gemma-4-31B-it architecture. Its primary objective is to integrate state-of-the-art reasoning capabilities, specifically distilled from Claude-4.6 Opus interactions. The model leverages datasets where reasoning effort was explicitly set to "High," enabling it to tackle complex problems with precise and nuanced solutions.

Key Enhancements (v2)

An enhanced version, gemma-4-31B-it-Claude-Opus-Distill-v2, offers significant improvements:

  • Dataset Quality: Re-curated high-density reasoning paths for superior response quality and logical depth.
  • Chat Template Fixes: Comprehensive structural fixes for improved formatting.
  • Generalization: Trained with a large batch size, high rank/alpha, and low learning rate to prioritize broad generalization.

Core Skills & Capabilities

Thanks to its robust base and high-effort reasoning distillation, this model is highly optimized for:

  • Coding: Advanced code generation, debugging, and software architecture planning.
  • Science: Deep scientific reasoning, hypothesis evaluation, and analytical problem-solving.
  • Deep Research: Navigating complex, multi-step research queries and synthesizing information.
  • General Purpose: Highly capable instruction-following for tasks requiring strong logical coherence.

Training Methodology

The model underwent Supervised Fine-Tuning (SFT) using Unsloth for efficient memory and compute optimization. It was trained extensively on reasoning trajectories from Claude Opus 4.6, utilizing high-quality datasets such as TeichAI/Claude-Opus-4.6-Reasoning-887x and Crownelius/Opus-4.6-Reasoning-2100x-formatted.