⚡ https://web.tribute.tg/d/KIH ⚡ If you like this Genesis LLM release you can donate to me via @Tribute bot in Telegram messenger and support future Genesis LLM development.

Gemma4-12B-QAT-Genesis

This is vanilla Gemma4-12B-QAT version from Google.

I distilled training and quantization noise from all tensors in it. Everything else left unchanged.

According to my tests she is perfect for roleplay, and works fast even on lowend GPUs.

Join the Discord for updates, roadmaps, projects, or just to chat.

⚡ Why Genesis project exists? During training, ALL models don't just learn knowledge - they also accumulate random noise in their tensors. This noise builds up and creates something I call the Noise Gate - a fundamental barrier that stops LLM models from learning further and makes them unstable, verbose, and prone to hallucinations. My approach reduces this noise. It repairs the signal without touching the learned knowledge and gradient. The result is a model that finally speaks clearly, follows instructions, and remembers context - because it's no longer fighting its own internal chaos.

~60% faster with MTP

Ships with an MTP (multi-token-prediction) draft head for speculative decoding — roughly 60% faster generation with identical output (the model verifies every drafted token, so quality is unchanged — pure speed). This release is tuned to pair well with the included MTP head.

llama.cpp:

llama-server \
  -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  -md mtp-gemma-4-12B-it.gguf --spec-type draft-mtp \
  -ngl 99 -fa on

Note: the MTP speedup was currently tested by me through llama.cpp (llama-server / llama-cli).

Downloads

Recommended settings from Unsloth

• temperature 1.0
• top_k 64
• top_p 0.95
• min_p disabled

Specs

• 12B dense · 256K (262144) context
• Vision (image input) via mmproj
• Based on Gemma 4 12B by Google DeepMind

Compatibility

• Works with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF runtimes.
• Multi-GPU + LM Studio: I've personally noticed Gemma 4 can crash under LM Studio's tensor-split mode — use a single GPU (layer-split or priority order) for this model.

Acknowledgements

• Google DeepMind — Gemma 4.
• The included mtp-gemma-4-12B-it.gguf speculative draft head comes from Unsloth's Gemma 4 release — many thanks to the Unsloth team for it.