An analytical encyclopedia for frontier model architectures. Compared, scored, and digested in real time — so you can see the field move before the press release lands.
Six axes, normalized. Value-Ratio dial recomputes reasoning-per-dollar on every swap. Click a slot to change the model.
Normalized 0–100 across all tracked systems.
Scroll the column on the left. The diagram on the right keeps pace, lighting up the part of the system the text is describing. Five stations, end to end.
Before anything can be reasoned about, your input is shattered into tokens — sub-word units drawn from a vocabulary of around 100k–200k pieces. "unforgettable" typically becomes un · forget · table.
Each token is looked up in an embedding table — a wide vector of learned numbers. From the model's point of view, your prompt now lives in a high-dimensional space where related ideas are geometrically close.
The transformer's defining trick: self-attention. Each token, in parallel, asks "which other tokens should I be paying attention to?" — and answers itself. Multiple attention heads ask this same question along different axes.
One head might track syntax, another follows entities across paragraphs, a third matches code brackets. The model isn't taught which head does what — that emerges from training.
An attention layer is paired with a feed-forward network — a dense MLP that reshapes the signal token-by-token. Stack 40–120 of these blocks and you have a transformer.
In mixture-of-experts models the feed-forward layer is replaced by many parallel experts; a router picks ~2 per token. That's how 671B-parameter DeepSeek runs with only 37B active per forward pass.
Pretraining — trillions of tokens of text and code. The model only ever learns "predict the next token", but at that scale, learning to predict becomes learning to compress, and compression starts to look like understanding.
Post-training — supervised fine-tuning on curated demonstrations, then preference optimisation (RLHF, DPO). This is where the helpful assistant persona appears.
RLVR — reinforcement learning from verifiable rewards. Used heavily for reasoning models: the model gets rewarded when its chain-of-thought leads to a correct, checkable answer.
What you call "GPT-5" or "Claude" in product is always a harness wrapped around model weights: a system prompt that defines its persona; a context window holding the conversation; a tool sheet listing functions it can call; a memory store; a safety filter on both sides.
Two products on the same weights can feel like different models. Two harnesses on the same task — agentic versus single-turn — can score 20 points apart on the same benchmark.
Tap a row to pull the drawer. Each dossier carries architecture, modality, and a live benchmark profile.
Pick a suite from the rail. Sort any column. Δ 30d shows movement since last month.