license: gemma library_name: gguf pipeline_tag: text-generation base_model: google/gemma-4-31b-it base_model_relation: finetune language:
- en
- ko tags:
- gemma
- gemma4
- gguf
- llama.cpp
- q4_k_m
- uncensored
- chat
- coding
- reasoning
SuperGemma4-31b-abliterated-GGUF
If this release helps you, support future drops on Ko-fi.
SuperGemma4-31b-abliterated-GGUF is the GGUF release of SuperGemma4-31b-abliterated, packaged for llama.cpp-compatible runtimes and built for people who want a fully uncensored, harder-hitting Gemma 4 31B experience on local hardware.
This release keeps the same product direction as the MLX version:
- fully uncensored chat with fewer brakes
- stronger coding and technical help
- sharper reasoning and planning
- better real-world usefulness for local users
- a surprisingly lightweight-feeling 31B deployment path for GGUF users
What you get
- GGUF quantized weights for local deployment
- Gemma chat template alongside the model files
- a straightforward path for llama.cpp, LM Studio, and other GGUF tooling
Why people will like it
This release was pushed toward the things end users notice immediately:
- much more open uncensored conversation
- stronger coding, debugging, and implementation help
- more useful answers on practical prompts instead of generic filler
- a local experience that feels sharper, more direct, and more builder-friendly
- a 31B release that feels leaner and punchier than the label suggests
In short: this is the Gemma 4 31B local drop for people who want fewer brakes, more capability, and a more addictive day-to-day experience.
Recommended usage
Example with llama.cpp:
llama-cli -m SuperGemma4-31b-abliterated.Q4_K_M.gguf -p "Write a clean FastAPI CRUD example." -n 256
Included clean-output helper
This release includes supergemma_guard.py and supergemma_gguf_guarded_generate.py for exact-output, JSON-only, and loop-sensitive workloads.
Example:
python supergemma_gguf_guarded_generate.py \
--model SuperGemma4-31b-abliterated.Q4_K_M.gguf \
--chat-template-file chat_template.jinja \
--prompt 'Return only valid JSON with keys "title" and "steps".'
Recommended behaviors:
- require raw JSON for JSON-only endpoints
- strip stray internal markers before presenting answers
- keep structured-output prompts explicit and narrow
- use the guarded runner by default for exact replies, fixed-line outputs, and tool-followup style prompts
Support
If you want to support more uncensored local model releases, benchmarks, and packaging work: