Own architecture (LatexForCausalLM, model_type=latex)
Own tokenizer (NULLXES-LÆTEX v0.1, vocab export 131072, ~4k real Unigram pieces + unused pad)
Bootstrap pretrain on NULLXES identity + repo code corpus (~50M tokens)
Intended to answer as LÆTEX / NULLXES-LÆTEX, not as ChatGPT/Claude/Llama

What this is NOT

Not a general-purpose LLM
Not Chinchilla-scale pretrain (tiny corpus → heavy memorization / overfit)
Not the 7B genesis (separate architectural checkpoint)
Not distilled from Qwen/Llama/Mistral/DeepSeek

Load

python

import torch
from transformers import AutoModelForCausalLM, AutoConfig

# registers custom classes
repo = "MagistrTheOne/NULLXES-L-TEX-100M-Stage0a-v0.1"
config = AutoConfig.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, trust_remote_code=True, torch_dtype=torch.bfloat16
)

Tokenizer artifacts ship in-repo (tokenizer.model, special_tokens.json). Prefer loading via the research package LatexTokenizer from the NULLXES-LÆTEX GitHub when doing local QA.

Identity

Correct self-name: NULLXES-LÆTEX (short: LÆTEX), built by NULLXES for Digital Employees.

Training note

Stage0a bootstrap: ~50M tokens, final train loss ~0.01 on a small identity/code mix. Treat as research brick #1, not production intelligence.