Blazed-Forge/Split-Untied-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

Blazed-Forge/Split-Untied-31B is a 31 billion parameter language model developed by Nimbz, built upon the Gemma 4 architecture. This experimental merge model, created using a three-phase StyleSwap merging technique, is specifically optimized for creative writing, darker themes, and enhanced lewdness, trading some instruction adherence for increased expressive freedom. It offers a distinct alternative to its counterpart, Split-31B, which prioritizes strict instruction following.

Loading preview...

Split Untied 31B: A Creative and Experimental Gemma 4 Merge

Split-Untied-31B is a 31 billion parameter model from Nimbz, based on the Gemma 4 architecture. It is an experimental merge that differentiates itself from its counterpart, Split-31B, by prioritizing creativity, darker themes, and lewdness over strict instruction adherence. This model is the result of a sophisticated three-phase merging process, utilizing two della_linear passes and a StyleSwap head graft.

Key Capabilities & Merging Process

The model's development involved integrating five different base models across three phases to achieve its unique characteristics:

  • Phase I: Comprehension focused on adjusting q_proj, k_proj, gate_proj, and up_proj parameters to influence persona tracking, world-state stability, format adherence, in-character responses, and swipe variety. Models like Melinoe-VL-heretic, Glistening-Gem v2.1, and MeroMero v2 contributed to this phase.
  • Phase II: Voice refined the model's expressive output by adjusting v_proj, o_proj, and mlp.down_proj parameters. This phase incorporated scotoma-2 for anti-Gemma tics, Dark-Thoughts V2 for a darker and less sanitized lean, Glistening-Gem v2.0 for self-repetition and coherence, and Melinoe-VL-heretic for willingness.
  • Phase III: The Growth (Head Graft) involved grafting an untied StyleTune lm_head from Gryphe/Pantheon-Reasoning-31B-1.1 onto the merged body, specifically at a t value of 0.85. This final step is responsible for the model's pronounced creative and 'darker' output.

Good For

  • Creative Writing & Roleplay: Excels in generating imaginative and less constrained narratives.
  • Darker & Lewd Content: Specifically tuned for outputs with a darker lean and explicit themes.
  • Experimental Use Cases: Ideal for users interested in exploring the boundaries of LLM capabilities and unique stylistic outputs.

Users seeking strict instruction adherence or highly structured outputs might prefer the Split-31B variant. GGUF quantizations are available from mradermacher for both static and imatrix versions.

Finetunes
1 models