RichWoollcott/architect-agent-gemma4-26b-moe

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 15, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

RichWoollcott/architect-agent-gemma4-26b-moe is a 26 billion parameter Mixture-of-Experts (MoE) model, fine-tuned from unsloth/gemma-4-26b-a4b-it. This model is specifically designed to emulate a senior software architect, applying Domain-Driven Design (DDD) strategic patterns and reasoning explicitly via internal thought blocks. It excels at assisting developers with architectural decisions, such as bounded-context design, integration patterns, and articulating trade-offs in complex systems, with a context length of 32768 tokens.

Loading preview...

Architect Agent (Gemma-4-26B-A4B MoE)

This model, developed by RichWoollcott, is a specialized fine-tune of the unsloth/gemma-4-26b-a4b-it 26 billion parameter Mixture-of-Experts (MoE) model. It is engineered to function as a senior software architect, focusing on applying Domain-Driven Design (DDD) strategic patterns and surfacing architectural tensions. A key characteristic is its explicit reasoning process, often presented in <think>...</think> blocks before providing a visible answer.

Key Capabilities

  • Architectural Decision Support: Assists with complex architectural challenges like bounded-context design, integration patterns, and sequencing legacy system migrations.
  • Trade-off Articulation: Capable of clearly articulating architectural trade-offs in written form.
  • Reasoning Transparency: Utilizes an internal reasoning block (<think>...</think>) to show its thought process, enhancing transparency and trustworthiness.
  • Integration: Designed for deployment within multi-agent systems, typically served via llama-swap, Ollama, or vLLM.

Training and Limitations

The model was fine-tuned using Unsloth + TRL on a single NVIDIA DGX Spark GB10, leveraging approximately 900 synthetic dialogues. These dialogues were generated by an adversarial Player-Coach pipeline, focusing on 10 architect-dimension categories. It's important to note that the model has no first-hand engineering experience and may confidently reason from incorrect premises if accepted. Therefore, it is recommended to pair it with a RAG (Retrieval Augmented Generation) layer for factual grounding from authoritative architecture sources. The training data is not publicly released due to license review, but the generator pipeline is open-source and reproducible.