glyphsoftware/gemma-4-26b-a4b-opus-4.7-distilled
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>thoughtreasoning 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.