Home / Atlas

LLM Atlas

An interactive map of how foundation models are built, adapted and served.

Explore the system from raw training data to the next generated token. Isolate subsystems, explode the architecture and inspect the mechanics behind Transformer layers, MoE routing, distributed training, post-training and high-throughput inference.

Built as a living systems map. Architecture details vary across model families and inference engines.

Seven subsystems are shown as plates carrying labelled components: data plane, model core, MoE subsystem, training, post-training, inference runtime and serving plane. Selecting a component opens its inspector.

Microscope

Implementation-level views

Each panel is a working simulation, not a screenshot. They run entirely in your browser.

Overview

Three phases, one system

01

Build

Data → tokenization → architecture → distributed pre-training.

02

Adapt

SFT → LoRA → preference optimization → evaluation.

03

Run

Prefill → KV state → scheduling → decoding → serving.

Sources

Sources & further reading

Primary papers and official documentation used for the technical facts in this atlas. Components cite them individually in the inspector.