SubMaroon/Boulesis-26B-A4B

Hugging Face
VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 4, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Warm

Boulesis-26B-A4B by SubMaroon is a 26 billion parameter composite Gemma 4 roleplay model with a 32768 token context length. It is designed to enhance narrative progression and character depth in roleplay scenarios by improving attention to context and diversifying prose. This model aims to move beyond passive reactivity, actively advancing the story while retaining its base intelligence, making it suitable for complex and action-oriented roleplay.

Loading preview...

Boulesis-26B-A4B Overview

Boulesis-26B-A4B is a 26 billion parameter composite Gemma 4 model developed by SubMaroon, specifically engineered for advanced roleplay (RP) scenarios. This model utilizes a unique blend of QK task arithmetic and fused LoRA to enhance its narrative capabilities without sacrificing core intelligence. The primary goal was to create a model that is more decisive, diversifies its prose, and exhibits sharper attention to context, allowing it to organically integrate lore and character details into the roleplay.

Key Capabilities

  • Proactive Narrative Advancement: Moves beyond simply reacting to user input, actively driving the story forward.
  • Enhanced Character Depth: Improves the model's ability to understand and utilize character cards, pulling relevant lore and facts into the roleplay.
  • Contextual Awareness: Designed to dig deeper into provided context, leading to more coherent and immersive interactions.
  • Zero Catastrophic Forgetting: Retains the base model's intelligence and knowledge while integrating specialized RP enhancements.
  • Composite Architecture: Built by selectively merging components (attention, lm_head, q_proj, k_proj, v_proj, o_proj) from various Gemma 4 fine-tunes to achieve specific behavioral traits.

Good for

  • Complex Roleplay Sessions: Especially effective for scenarios involving multiple characters and intricate plotlines due to its improved memory and character grasp.
  • Action-Oriented Narratives: The v1 version is noted for delivering more dynamic plots and driving characters to take action.
  • Detailed Character Interactions: Excels at maintaining character consistency and integrating background information naturally.
  • Users seeking decisive and engaging AI partners: The model is designed to be less passive and more active in co-creating the narrative.

Finetunes
1 models