Nimbz/Split-31B

Hugging Face
VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Nimbz/Split-31B is an experimental 31 billion parameter language model based on the Gemma 4 architecture, created by Nimbz through a two-phase della_linear merge process. This model is specifically optimized for strong instruction adherence, persona tracking, and maintaining world-state stability over long sessions, making it suitable for structured output requirements. It offers a context length of 32768 tokens and is designed for roleplay scenarios requiring consistent character and format adherence.

Loading preview...

Nimbz/Split-31B: A Dual-Personality Gemma 4 Merge

Nimbz/Split-31B is an experimental 31 billion parameter language model built upon the Gemma 4 architecture, developed by Nimbz using a two-phase della_linear merge method. This model is one of two distinct 'heads' from the same core, each optimized for different output characteristics.

Key Capabilities & Design Philosophy

Split-31B is engineered for strong instruction adherence, persona tracking, and maintaining world-state stability throughout extended conversational sessions. Its development involved carefully adjusting specific layers of the underlying Gemma 4 models:

  • Phase 1 (Comprehension): Focused on modifying q_proj, k_proj (for attention mechanism) and gate_proj, up_proj (for knowledge routing) from three base models: Melinoe-VL-heretic (for persona tracking and world-state), Glistening-Gem v2.1 (for format adherence and in-character responses), and MeroMero v2 (for output variety).
  • Phase 2 (Voice): Adjusted v_proj (for expression) and o_proj, mlp.down_proj (for output writing) using scotoma-2 (to counter Gemma's inherent 'tics'), Dark-Thoughts V2 (for a 'darker lean' and less sanitization), Glistening-Gem v2.0 (for self-repetition and coherence), and Melinoe-VL-heretic (for 'willingness').

When to Use Split-31B

Choose Nimbz/Split-31B if your application requires:

  • Strict instruction following: Ideal for tasks demanding precise formatting, such as trackers or specific header/footer blocks.
  • Consistent character and persona: Excels in roleplay scenarios where maintaining a stable persona and world-state is crucial.
  • Long-session coherence: Designed to stay on track and provide structured output over extended interactions.

For use cases requiring more creativity, a 'darker lean,' or 'lewdness' at the expense of some prompt adherence, the companion model, Split-Untied-31B, is recommended. The model supports a context length of 32768 tokens.

Finetunes
1 models