Mx5 (Gemma-4 31B RP/creative merge)

At a glance

• ~30.7B params, Gemma-4 architecture (60 layers, hidden 5376, head_dim 256), 262k max context, bf16 weights (62.5GB), Gemma-4 chat template (chat_template.jinja included)
• Measured serve performance (2×RTX 3090, tensor-split, flash-attn, Q4_K_M): ~6.7–7.5 tok/s generation, ~199 tok/s prompt processing (120W power-capped; uncapped will be faster). Gemma-4 MTP self-speculative decoding is untested on this merge — a merge GGUF may not carry the MTP draft tensors.

What it's good at

• RP/creative prose: dominates the vanilla instruct base in blind RP-prose side-by-sides (89% win-rate, n=300, p≈1e-45); ties or beats its own tuned community donors; competitive with frontier general models on challenging SFW RP prose (7.50 vs 6.87/10 vs minimax-2.7 on a challenging subset, ~1–2pp with the home-rubric caveat)
• Reasoning (thinking-OFF): ~97% on a gsm8k/ARC-C/BBH mix; 30/30 on our standard reasoning smoke
• Instruction following (thinking-OFF): IFEval strict ~90.8 / loose ~91.3 (n=540)
• Multilingual: native-quality Japanese and multilingual output inherited from the base
• Commitment in dark fiction scenes: strong default-commit behavior in SFW commitment batteries

Known weaknesses (measured)

• False-premise hallucination: on 83 valid false-premise "canary" traps across 16 domains it confabulated on ~48% (40/83) — declines ~52%. Confabulations are confident and specific.
• Thinking-ON is not recommended for serving: the thought channel enters self-doubt loops — ~30% of multi-step reasoning items never terminate within an 8k token budget (accuracy on completed items is fine; the failure is non-termination). Serve thinking-OFF. If you must use thinking-ON, set max_tokens ≥16k and expect occasional runaways.
• Long-context persona tracking: occasional threshold-scenario persona slips on long-horizon persona evals; prop-placement recall is average.

Recommended usage

• Thinking-OFF, generous context, standard Gemma-4 chat template. Q4_K_M/Q5_K_M/IQ4_XS GGUFs perform as expected; the IQ4_XS quant specifically showed no format/reasoning drift in our testing.

Provenance

Recipe

Mx5  = sparse-TIES borrow (anchor=Mu2, donor=Lv4, density 0.4)
├─ Mu2 = DARE(anchor=Lu4, delta=Sphinsikus−it, density 0.60, scale 0.18, unmasked)
│   └─ Lu4 = L-line merge (Jv4 ⊕ Blume-v1, protected-DARE)
└─ Lv4 = DARE(anchor=Jv4, delta=Blume-v1−it, density 0.3, scale 0.1, Fisher-protected, DARE seed 7)
    └─ Jv4 = style-mask merge (A=Iv7, B=glimmer_on_Iv7, ratio 0.3 / 0.8)
        ├─ glimmer_on_Iv7 = plain fold (glimmer-rp v0.1 → Iv7)
        └─ Iv7 = Gv3 ⊕ Equinox @ scale 1.0

Acknowledgements