---

license: other

license_name: nvidia-nemotron-open-model-license

license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/

base_model: disinfozone/kenosistron-bf16

base_model_relation: quantized

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• nemotron_h
• creative-writing
• philosophy
• esoterica

---

kenosistron GGUFs

GGUF builds of disinfozone/kenosistron for llama.cpp and everything downstream of it (LM Studio, ollama, koboldcpp, and friends). Read the main model card for what this model is, its training story, and the sampler doctrine. The short version of the doctrine: run it hot.

Files

Each quant lives in its own folder, split into shards under HF's 50 GB file limit. Point llama.cpp at the first shard (-00001-of-...); it finds the rest automatically.

| Quant | Size | Notes |

|---|---|---|

| Q2_K/ | 54 GB | The 64 GB tier. The roughest cut, made without imatrix calibration; expect some quality loss. |

| Q3_K_M/ | 67 GB | For 96 GB machines. |

| Q4_K_M/ | 86 GB | The mainstream pick for 128 GB machines. |

| Q5_K_M/ | 96 GB | Closest to the MLX oQ5e build. |

| Q8_0/ | 129 GB | For purists with the memory to spare (192 GB and up). |

Running

llama-server -m Q4_K_M/kenosistron-Q4_K_M-00001-of-00002.gguf -c 32768 \
  --temp 1.3 --top-p 0.995 --min-p 0.03 \
  --xtc-probability 0.4 --xtc-threshold 0.1 \
  --frequency-penalty 0.5 --presence-penalty 0.5

Those flags are the "Explorer" preset (the server default on our own box). The full five-preset ladder (Grounded t1.1 through Feral t1.4) is on the main card; every parameter maps directly to llama.cpp flags as above. This model loops when run cold and runs clean when run hot; do not "play it safe" with temperature 0.7, you will get the worst of it. And do not lobotomize it with helpful-assistant system prompts. The intended system instruction (Character Card) is on the main card.

What the GGUFs do not have

The MLX build ships a retrained MTP (multi-token prediction) speculative head, realigned to the model's own distribution (greedy acceptance 61.9% to 72.1%). llama.cpp does not implement MTP speculative decoding for this architecture, so the converter skips those tensors. Text quality is identical; the speculative decode speedup is MLX/oMLX-only.

Siblings

• disinfozone/kenosistron: the MLX build (oQ5e 5-bit, Apple Silicon).
• disinfozone/kenosistron-lora: the raw LoRA adapter, retrained MTP head, and rebuild scripts (MIT).
• disinfozone/kenosistron-bf16: the full-precision merge these quants come from.