appbinary-inspector

Static inspection tools plus an open-source, disk-paged character-model architecture.

Features

• Detects PE, ELF, Mach-O, ZIP, 7z, and RAR headers
• Calculates SHA-256 and SHA-1 hashes
• Extracts readable strings and entropy
• Reads basic PE header metadata
• Extracts embedded ZIP archives and carves recognizable binary signatures
• Exports PE headers and raw binary sections for authorized analysis
• Builds a compact tokenizer for source code, logs, or extracted text
• Provides editable PyTorch classifier, autoencoder, and transformer templates
• Keeps CardStack model cards as separately stored 4-bit SSD files
• Loads only router-selected cards into a small LRU cache during generation
• Generates one next byte at a time for true character-level output

Install

pip install appbinary-inspector

Use

appbin inspect program.exe
appbin extract program.exe --output extracted
appbin binary-extract program.exe --output pe_parts
appbin tokenize source.py --vocab source-tokenizer.json
appbin model-template classifier

CardStack: low-VRAM model experiments

CardStack is a small trainable byte-level language model. Its expert cards are stored as separate 4-bit files. At inference, inactive cards remain on disk and the cache retains only a limited number of recently selected cards.

from appbinary_inspector import CardStackConfig, CardStackLM, train_next_character

model = CardStackLM(CardStackConfig(width=128, layers=4, cards_per_layer=8, active_cards=2, cache_cards=4))
losses = train_next_character(model, open("training.txt", encoding="utf-8").read(), steps=500)
model.save("my-card-model")  # writes a base file plus separate 4-bit card files

paged = CardStackLM.load("my-card-model")
print(paged.generate("Hello", new_characters=80))

CLI version:

appbin card-create --output my-card-model --width 128 --layers 4 --cards 8 --active-cards 2
appbin card-train my-card-model training.txt --steps 500
appbin card-generate my-card-model "Hello" --characters 80

Text chat interface

Open a local chat box after training a model. It binds only to your computer, opens in your browser, and never sends prompts to a remote service.

appbin chat my-card-model

Python import API

Use the model from normal Python code without a terminal command interface:

from appbinary_inspector import appbincardcr

mycard = appbincardcr("my-card-model")
reply = mycard("Explain gravity", output=True)

# Or call it in one line:
reply = appbincardcr(mycard, "Explain gravity", output=True)

Hugging Face models

Hub IDs and local Transformers directories work through the same API.

from appbinary_inspector import appbincardcr

mycard = appbincardcr(
    "Qwen/Qwen2.5-1.5B-Instruct",
    backend="huggingface",
    device_map="auto",
)
mycard("Explain gravity", output=True)

Install support with pip install "appbinary-inspector[huggingface]". For CUDA 4-bit loading, install appbinary-inspector[huggingface-4bit] and pass load_in_4bit=True. Standard Hugging Face models keep their original dense layout; only models trained or converted for CardStack use disk-card paging.

This reduces active model-card memory. It does not make a dense, unmodified 70B model fit instantly on a low-VRAM GPU: disk bandwidth and CPU/GPU transfer time become the tradeoff, and models need to be trained for sparse routing.

Convert a large model for low VRAM

appbin quantize creates a 4-bit GGUF for compatible locally downloaded Hugging Face model folders. GGUF runs in llama.cpp and apps such as LM Studio, where CPU and SSD offload can be used instead of requiring large VRAM.

Install a current llama.cpp build first. The command needs its convert_hf_to_gguf.py script and llama-quantize executable:

appbin quantize C:\\models\\DeepSeek-R1-Distill-Qwen-32B --output C:\\models\\deepseek-32b-q4.gguf --level Q4_K_M --converter C:\\llama.cpp\\convert_hf_to_gguf.py --llama-quantize C:\\llama.cpp\\build\\bin\\Release\\llama-quantize.exe

Q4_K_M is the recommended start. A 32B Q4 model still needs roughly 18–22 GB of disk space and substantial system RAM, but does not require a large GPU. Conversion temporarily needs extra disk space for an F16 GGUF.

PyTorch models

Install optional PyTorch support:

pip install "appbinary-inspector[torch]"

from appbinary_inspector.torch_builder import create_model, preset

config = preset("classifier")
config["layers"][0]["out_features"] = 512
model = create_model(config)

Safety

Targets are read as bytes only and never executed. The package does not defeat encryption, DRM, licensing controls, passwords, or code protections. It can analyze plaintext, export raw PE sections, and extract recognizable embedded content from files you are authorized to inspect.