chaiops/llm-arch

Llm-arch is a terminal based tool that lets you explore the internal architecture of a large language model without ever loading the actual model weights into memory. It works by reading only the header information from model checkpoint files, in formats called safetensors and GGUF, which describes the shape and structure of the model rather than its full numerical content. Because of this, even a huge model with billions of parameters opens almost instantly, since the heavy data itself is never touched. You run the tool from the command line by pointing it at a model file, a folder containing a sharded model split across several files, an index file describing those shards, or a GGUF format model used by tools like llama.cpp. Once open, it displays a summary of the model, including its architecture family, total parameter count, size on disk, number of layers, hidden size, number of attention heads, vocabulary size, and data types, reading this from a configuration file when one is available and otherwise working it out from the shapes of the model's internal tensors. The main view presents the model as a visual stack, with each part of the model, such as the embeddings, each transformer layer, the final normalization step, and the output head, drawn as its own block connected to the others by the flow of information through the network. Selecting a transformer layer expands it to show its internal structure, like the attention and feed forward components, each with its own parameter count. A second view shows the model's tensors as labeled rectangular blocks with their real dimensions, tracing how data actually flows through a transformer block from input to output. A left hand panel lists every layer as an expandable tree with parameter counts, and a right hand panel shows detailed information about whatever tensor or group is currently selected, including its role within the architecture. The tool is navigated entirely with the keyboard, using simple keys to move around, expand or collapse sections, filter by tensor name, jump to the top or bottom, and quit. It is written in Go, can be installed with a single shell command or through Go's own package installer, and currently has 6 stars on GitHub, aimed at people who want to understand how a specific language model is built without downloading or running the full model.