WaveCut/Gemma-4-96E-A4B-Heretic-TQ

EXPERIMENTAL, AVOID DEPLOYING TO SENSITIVE ENVIRONMENTS

---

Gemma-4-96E-A4B-Heretic-TQ GGUF

TurboQuant GGUF builds of blascotobasco/Gemma-4-96E-A4B-Heretic for the TurboQuant llama.cpp branch.

These files were refreshed on 2026-04-08 so that the embedded tokenizer.chat_template matches the updated Gemma 4 Interleaved template. The published TQ GGUFs now work in reasoning mode on the linked branch without requiring --chat-template-file.

Files

File	Size	Notes
Gemma-4-96E-A4B-Heretic-TQ3_1S.gguf	12589895200 bytes (11.73 GiB)	TQ3_1S, updated embedded interleaved template
Gemma-4-96E-A4B-Heretic-TQ4_1S.gguf	13985279520 bytes (13.02 GiB)	TQ4_1S, updated embedded interleaved template
chat_template.jinja	standalone Gemma 4 Interleaved template matching the embedded GGUF template
requant_recipe_tq3_1s.txt	tensor override recipe for TQ3_1S
requant_recipe_tq4_1s.txt	tensor override recipe for TQ4_1S

Runtime

Use this exact llama.cpp branch:

• https://github.com/iamwavecut/llama-cpp-turboquant/tree/feature/turboquant-kv-cache

Related project documentation:

• https://github.com/TheTom/turboquant_plus/tree/main

Build explicitly from that branch:

git clone https://github.com/iamwavecut/llama-cpp-turboquant.git
cd llama-cpp-turboquant
git checkout feature/turboquant-kv-cache
cmake -S . -B build
cmake --build build --config Release -j

Stock upstream llama.cpp is not the supported runtime for these TurboQuant GGUFs.

Template and reasoning

• The default embedded template is Gemma 4 Interleaved.
• chat_template.jinja in this repo matches the embedded template.
• The 2026-04-08 refresh adds a safe fallback system message when --reasoning on is enabled and no explicit -sys prompt is provided.
• Because of that change, --chat-template-file is no longer required for normal reasoning usage on the supported branch.
• Tensor data was not requantized again for this refresh; the published files were repacked to update GGUF metadata and embedded template content.

Recommended launch commands

TQ4_1S

Reasoning enabled:

/path/to/llama-cli \
  -m /path/to/Gemma-4-96E-A4B-Heretic-TQ4_1S.gguf \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv turbo4 \
  -c 8192 \
  -cnv \
  --jinja \
  --reasoning on

Reasoning disabled:

/path/to/llama-cli \
  -m /path/to/Gemma-4-96E-A4B-Heretic-TQ4_1S.gguf \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv turbo4 \
  -c 8192 \
  -cnv \
  --jinja \
  --reasoning off \
  --reasoning-format none

TQ3_1S

Reasoning enabled:

/path/to/llama-cli \
  -m /path/to/Gemma-4-96E-A4B-Heretic-TQ3_1S.gguf \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv turbo3 \
  -c 8192 \
  -cnv \
  --jinja \
  --reasoning on

Reasoning disabled:

/path/to/llama-cli \
  -m /path/to/Gemma-4-96E-A4B-Heretic-TQ3_1S.gguf \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv turbo3 \
  -c 8192 \
  -cnv \
  --jinja \
  --reasoning off \
  --reasoning-format none

Notes

• If you already downloaded an older copy of these TQ files and reasoning mode failed without an external template override, download the refreshed files from this repo.
• Q8_0 is the clean repacked source checkpoint used locally to produce the published TurboQuant requants, but it is not distributed from this repo.
• SQ variants are intentionally not published here.

Credits

• Base model author: blascotobasco
• TurboQuant runtime / GGUF work based on llama.cpp and the linked TurboQuant branch