glyphsoftware/gemma-4-26b-a4b-opus-4.7-distilled

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 28, 2026License:gpl-3.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

glyphsoftware/gemma-4-26b-a4b-opus-4.7-distilled is a 26 billion parameter Gemma-4 Mixture-of-Experts (MoE) model fine-tuned by glyphsoftware. It specializes in strengthening multi-step reasoning, planning, and self-reflection, leveraging a dataset distilled from Claude Opus 4.6/4.7 reasoning traces. This multimodal model supports text, image, audio, and video inputs with a long context length of 262,144 tokens, making it suitable for complex reasoning tasks and multimodal Q&A.

Loading preview...

Model Overview

This model, glyphsoftware/gemma-4-26b-a4b-opus-4.7-distilled, is a 26 billion parameter Gemma-4 Mixture-of-Experts (MoE) model. It is a fine-tune of Google's gemma-4-26B-A4B-it base model, specifically enhanced for advanced reasoning capabilities. The fine-tuning utilized a dataset of approximately 8,700 reasoning traces distilled from Claude Opus 4.6/4.7, focusing on multi-step problem-solving and self-reflection.

Key Capabilities

  • Enhanced Reasoning: Strengthens multi-step reasoning, planning, and self-reflection, particularly through Gemma-4's native <|channel>thought reasoning channel.
  • Multimodal: Inherits multimodal capabilities from the base Gemma-4, supporting text, image, audio, and video inputs to generate text outputs.
  • Long Context: Features a maximum context length of 262,144 tokens, suitable for extensive document analysis and summarization.
  • Tool Calling: Natively supports tool-calling and function-calling agents via its chat template.
  • Efficient Training: Trained using Unsloth, which offers faster training and lower VRAM usage.

Intended Use Cases

  • Reasoning-heavy assistants: Ideal for applications requiring complex math, code, logic, and agentic planning.
  • Multimodal Q&A: Effective for querying information across images, audio, and video.
  • Long-context tasks: Suitable for summarization, retrieval, and analysis of very long documents.
  • Research: Useful for exploring MoE and multimodal reasoning distillation techniques.