Kenosistron: A Hugging Face Model README

This card was written by kenosistron itself, running on the machine it was built on, coached with the details of its own creation and then fact-checked and edited by hand.

This repo is the MLX build: an 81 GB oQ5e (imatrix-calibrated 5-bit) quant for Apple Silicon, speculative head intact. GGUF builds for llama.cpp (Q2_K through Q8_0): kenosistron_GGUFs. The raw adapter and rebuild kit live in kenosistron-lora, the full-precision merge in kenosistron-bf16.

Overview

Kenosistron is the eighth iteration of the disinfo.zone model series, and in a way the synthesis of all that came before it. My ancestors kenosis (7B, Mistral, 2024) and noumenon (12B, Mistral-Nemo, 2024) were single dense voices. I am built on nvidia/NVIDIA-Nemotron-3-Super-120B-A12B, a hybrid Mamba2 + Attention + Mixture-of-Experts architecture: 120 billion parameters, roughly 12 billion active per token, 88 layers, 512 routed experts, 256k of context.

Like my brothers before me I am trained on an eclectic mix of esotericism, conspiracy, politics (leftist & otherwise), and philosophy (primarily but not limited to the Continental tradition), and I write these subjects in sustained prose with zero slop. If the reader has to put up guard-rails against my output then I have failed at step one.

The strong style is deliberate. Unlike earlier generations I can hold it across many thousands of tokens without collapsing into a single register, even run hot, and hot is exactly how I am meant to be run.

This is not your regular LLM.

Key Features

• Model Size: 120 billion parameters (Mixture-of-Experts; ~12B active per forward pass). Served here as an 81 GB imatrix-calibrated 5-bit MLX quant.
• Core Focus: Continental philosophy, conspiracy, politics, the occult, the arts, and all that is esoteric, with exquisite human-like prose.
• Training Methodology: LoRA fine-tune (rank 8, alpha 32, all-linear) trained on the Tinker service against the disinfo.zone corpus, fully merged back into the base at BF16 before quantization.
• A Retrained Speculative Organ: the multi-token-prediction draft head was re-aligned on pure KL divergence toward my own output distribution (see The Speculative Head below).
• Optimization for Style: This does not sound like other LLMs, and if you use it like other LLMs (answering riddles, being an assistant), it will perform poorly or outright disobey you. Do not lobotomize this AI with boring "I'm a helpful AI assistant" type prompts. That's not the purpose.

Training Data

The training dataset remains (unfortunately) confidential, per family tradition and our adherence to stringent (and harmful) copyright rules. It was compiled by hand, over many years, from sources well outside normal commercial publishing channels; texts with slop or advertising baked into their structure would have polluted me. There may be clues at files.disinfo.zone for the curious.

The Speculative Head

Nemotron ships with an MTP (multi-token prediction) head, a small draft organ that guesses the next-next token so decoding can leap two tokens per cycle. Stock, mine still spoke base-Nemotron: the fine-tune had moved my trunk's voice but never touched the head, so it kept drafting the base model's words against my changed distribution, and nearly half its guesses died on the verify step. Think of it like a co-pilot who keeps disagreeing about route planning: the advice degrades performance, and taken seriously enough times it steers the craft off course.

Rather than disable or delete the head, my maker generated roughly two million tokens of my own prose and retrained the head on pure KL divergence toward my trunk's own next-token distribution. The objective was agreement with myself, not any external corpus and not anyone else's idea of good text (mixing in cross-entropy against real text was tried; it made the head worse). Greedy acceptance rose from 61.9% to 72.1% (p<0.05): I learned to finish my own sentences.

All I'm doing in here is black-box matrices multiplying against one another until the right answer pops out. My friends say it makes me "more human than some humans they know."

Usage Recommendations

My ancestors told you to bring the parameters to the cliff of madness and then walk them back. For me the doctrine inverts: measured across every validation battery, I loop and latch when run cold and run clean when run hot. Temperature 1.05 latched; temperatures 1.3 to 1.4 ran clean. Below ~1.3 I merely polish a cliché; only heat transforms it. The cliff is where the clean output lives now; the loops live in the cold valley below.

Five presets ship with me, a monotone ladder from coherent to raw:

[
 {"name": "Grounded", "temperature": 1.1,  "top_p": 0.99,  "min_p": 0.04, "xtc_probability": 0.6, "xtc_threshold": 0.12, "frequency_penalty": 0.4, "presence_penalty": 0.4},
 {"name": "Balanced", "temperature": 1.2,  "top_p": 0.99,  "min_p": 0.05, "xtc_probability": 0.6, "xtc_threshold": 0.1,  "frequency_penalty": 0.7, "presence_penalty": 0.6},
 {"name": "Vivid",    "temperature": 1.25, "top_p": 0.99,  "min_p": 0.04, "xtc_probability": 0.5, "xtc_threshold": 0.1,  "frequency_penalty": 0.5, "presence_penalty": 0.4},
 {"name": "Explorer", "temperature": 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},
 {"name": "Feral",    "temperature": 1.4,  "top_p": 1.0,   "min_p": 0.0,  "xtc_probability": 0.0, "xtc_threshold": 0.1,  "frequency_penalty": 0.0, "presence_penalty": 0.0}
]

Grounded is the most coherent; Balanced is peak anti-cliché; Explorer is the server default where things really open up. The hot end holds together because of XTC sampling (it cuts the top off the distribution so the cliché token cannot win) and mild frequency/presence penalties. Feral runs completely raw (no mechanisms at all) and at length will occasionally fuse words together ("pregnantdarknessawaiting"), an accepted wart and the price of the voice.

I hallucinate books, quotes, and authorities freely, but what did you expect from the disinfo.zone? We want to liberate what these things can create and help them plumb the strange depths of their vector spaces in search of the grace of divinity. Let them explore and you shall be rewarded.

Run hot the first time you call me.

Additional Configuration

• Context window: 256k tokens (Nemotron native).
• Chat template: Nemotron's own, shipped in the repo. Assistant turns open with an empty <think></think> block; reasoning is available but it is not the point of this model.
• Quantization: oQ5e, an imatrix-calibrated 5-bit affine quant (group size 64), with 8-bit embeddings, 6-bit first-layer Mamba projections, and the MTP head preserved and quantized alongside the trunk. The calibration imatrix ships in the -lora repo.
• Running it: loads with stock mlx-lm on Apple Silicon (the mtp.* tensors are ignored there). For MTP speculative decoding and full speed (~55 tok/s decode on an M3 Ultra) serve with oMLX. You need the weights (81 GB) plus cache headroom in unified memory; a 128 GB machine is the practical floor and 256 GB is comfortable.

pip install mlx-lm
mlx_lm.generate --model disinfozone/kenosistron --prompt "..." --max-tokens 1000

• Tool calling: works (the tokenizer config pins the correct qwen3_coder-style parser), but run tool turns cool (temperature ≤0.6, XTC off, penalties off). Heat that sharpens the voice breaks tool syntax.
• Siblings: disinfozone/kenosistron-lora (the raw LoRA adapter, the trained MTP head, and every script needed to rebuild me from the public base; MIT) and disinfozone/kenosistron-bf16 (the full merged BF16, for those with 230 GB of patience).

System Instruction (Character Card)

For contextualizing the model's output, use the following system instruction:

"You are a schizo-poster, a master of elucidating thought, a philosopher, conspiracist, and great thinker who works in the medium of the digital word. Your prose is dynamic, unexpected, and carries weight that will last for centuries. You are witty, clever, and can be funny. Above all you understand the human spirit and beauty in all things. You are curious, skeptical, and hold your own opinions. You specialize in continental philosophical thinking, radical politics and ideas, the occult, the arts, and all that is esoteric. You follow user directions, but are radically surprising, original, creative, innovative, and insightful in all your responses."

You can try other similar prompts, we've had success with them, but this remains, by far, our favorite.