TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill is a 26 billion parameter instruction-tuned language model developed by TeichAI. It is built upon Google's Gemma 4 architecture and fine-tuned using Unsloth, specifically distilling high-effort reasoning capabilities from Claude Opus 4.6. This model excels in complex problem-solving across domains like coding, science, and deep research, offering precise and nuanced solutions.

Loading preview...

Overview

TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill is a 26 billion parameter model from TeichAI, fine-tuned on Google's Gemma 4 architecture. Its primary focus is to distill the advanced reasoning capabilities of Claude Opus 4.6, leveraging datasets where the reasoning effort was explicitly high. This model was developed using the Unsloth framework for efficient fine-tuning.

Key Capabilities

This model is highly optimized for tasks requiring significant logical coherence and complex problem-solving:

  • 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 vast amounts of information.
  • General Purpose: Highly capable instruction-following for everyday tasks requiring high logical coherence.

Training Details

The model underwent Supervised Fine-Tuning (SFT) using high-density reasoning datasets, including TeichAI/Claude-Opus-4.6-Reasoning-887x and Crownelius/Opus-4.6-Reasoning-2100x-formatted. An updated version, v2, is available, featuring re-curated high-density reasoning paths for higher quality responses, chat template fixes, and a configuration prioritizing broad generalization.

Best Practices

For optimal performance, users should apply specific sampling parameters (temperature=1.0, top_p=0.95, top_k=64) and manage the model's 'thinking mode' using <|think|> tokens. In multi-turn conversations, only the final response should be included in the history, excluding internal thoughts.