CiscoKpanse/sp-DNM-4-26B-A4B-it_v0.1

VISIONConcurrency Cost:2Model Size:26BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 20, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

CiscoKpanse/sp-DNM-4-26B-A4B-it_v0.1 is a 26 billion parameter instruction-tuned multimodal Mixture-of-Experts (MoE) model from the Google DeepMind Gemma 4 family, featuring a 256K token context window. This model processes text, image, and video inputs, generating text outputs, and is optimized for reasoning, coding, and agentic capabilities. Its MoE architecture allows for efficient inference, running with 3.8 billion active parameters, making it suitable for consumer GPUs and workstations.

Loading preview...

Gemma 4 26B A4B MoE: Multimodal Reasoning and Efficient Inference

This model, CiscoKpanse/sp-DNM-4-26B-A4B-it_v0.1, is a 26 billion parameter instruction-tuned variant from Google DeepMind's Gemma 4 family. It is a multimodal Mixture-of-Experts (MoE) model, capable of processing text, image, and video inputs to generate text outputs. A key differentiator is its MoE architecture, which, despite having 25.2 billion total parameters, only activates 3.8 billion parameters during inference, offering significantly faster performance comparable to a 4B model.

Key Capabilities

  • Multimodality: Handles text, image, and video inputs with variable aspect ratio and resolution support. It supports interleaved multimodal input, allowing free mixing of text and images in prompts.
  • Reasoning: Designed with configurable thinking modes and native system prompt support for structured conversations.
  • Extended Context: Features a substantial 256K token context window.
  • Coding & Agentic Capabilities: Shows enhanced performance in coding benchmarks and includes native function-calling support for autonomous agents.
  • Multilingual Support: Pre-trained on over 140 languages with out-of-the-box support for 35+ languages.

Good for

  • Efficient Inference: Ideal for applications requiring high performance on consumer GPUs and workstations due to its MoE design.
  • Complex Multimodal Tasks: Excels in scenarios involving image understanding (object detection, OCR, document parsing), video analysis, and combining various input types.
  • Reasoning and Code Generation: Strong performance in reasoning tasks and coding benchmarks, making it suitable for development and problem-solving applications.
  • Agentic Workflows: Native function-calling support facilitates the creation of highly capable AI agents.

Finetunes
1 models