GemStrike 31B — EXL3 4.00 bpw

GemStrike is a Gemma 4 31B creative-writing and roleplay fine-tune. This release is the EXL3 4.00 bpw conversion of the merged BF16 model.

The model was produced by training a language-only LoRA on the BF16 safetensors release of Huihui Gemma 4 31B IT QAT unquantized abliterated, merging that adapter into the BF16 base, and converting the merged checkpoint with ExLlamaV3.

Format

Quantization: EXL3, 4.00 target bits per weight
Language-model head: 6-bit
Codebook: mul1
Vision tower: preserved in BF16
Calibration: 1,407 rows × 2,048 tokens from a deterministic 512-example creative-writing/roleplay calibration selection
Storage: three safetensors shards

This is not a GGUF model. Use an EXL3-compatible ExLlamaV3 runtime, such as ExLlamaV3 itself or a compatible serving frontend. It is not intended for llama.cpp.

Fine-tune details

Training examples: 4,546
Held-out validation examples: 264
Final optimizer step: 4,546
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0
Adapter targets: language-model attention and MLP projections only
Vision and other non-language components were frozen during LoRA training

The training corpus was the ordered creative-writing/roleplay split used for the associated GemStrike training run.

Runtime validation

This published directory passed the following local acceptance checks:

Complete EXL3 tensor and safetensors-index coverage
411 EXL3 linear modules
All 355 vision tensors present in BF16
Deterministic text generation with ExLlamaV3 using two-GPU tensor parallelism
Successful unload and GPU-memory return-to-baseline checks

The runtime generation check used a 4,096-token cache and a fixed held-out text prompt. The preserved vision tower was audited statically, but image-input generation was not separately validated as part of this publication gate.

Prompting

Use the included Gemma 4 chat template and tokenizer metadata. For best results, send structured chat messages rather than manually inventing control tokens.

This fine-tune emphasizes creative prose and roleplay. It can produce mature, controversial, or otherwise sensitive material, especially because its base model is abliterated. Applications should apply their own moderation and review appropriate to their audience and use case.

License and attribution

Use is subject to the Gemma Terms of Use. This derivative also retains the lineage and conditions of its Huihui base model. Review those terms before redistribution or deployment.