NEXUS AI
Compiling Architecture · 000
Technology / Under the Hood

How NEXUS
thinks.

Scroll to watch the architecture assemble itself, block by block. Every concept you scroll past becomes visible — and active — in the diagram.

Input Tokens
Step 01 Embeddings + Positions
Step 02 Multi-Head Attention
Step 03 Feed-Forward Network
Step 04 × 96 layers, residual + norm
Step 05 Output Distribution
[ 00 ] Input

Begin with tokens.

Every input — text, image patches, audio samples — is broken into tokens, the atoms of meaning. NEXUS speaks 134 languages, but the model only knows tokens.

[ 01 ] Embed

Project into latent space.

Each token becomes a vector in a 12,288-dimensional space. Position encodings tell the model where each token sits in the sequence — order matters.

[ 02 ] Attend

Multi-head attention.

Every token looks at every other token, in parallel, across 96 attention heads. This is the trick that lets the model reason about long contexts — and it scales gracefully.

def attention(q, k, v):
  scores = q @ k.T / sqrt(d_k)
  weights = softmax(scores)
  return weights @ v
[ 03 ] Transform

Feed-forward refinement.

Two dense layers per block sharpen the representation, projecting through a 49,152-dim hidden space. This is where most of the parameters live, and where most of the "knowledge" is stored.

[ 04 ] Stack

Ninety-six times deep.

The attention + FFN block repeats 96 times, with residual connections and layer normalization between each. Information flows up, gets refined, flows up again.

[ 05 ] Emit

A probability distribution.

The final layer projects back to vocabulary space — a probability for every possible next token. We sample. We continue. A response is born, one token at a time, at 240 tokens/sec.

[ 06 ] Training

A million GPU hours, distilled.

NEXUS is trained in four stages — each stage compounding the previous one's gains.

STAGE 01 — PRE-TRAINING

A diet of internet.

Trillions of tokens — books, code, scientific papers, the open web. The model learns the texture of language.

STAGE 02 — FINE-TUNING

Curated examples.

High-quality demonstrations from experts. The model learns how to be helpful, not just what humans write.

STAGE 03 — RLHF

Human preference.

Reinforcement learning from human feedback. Humans rank outputs; the model learns the ranking. Subtle, hard, essential.

STAGE 04 — CONSTITUTION

Self-critique.

The model critiques its own outputs against a constitution of principles. Alignment without endless human labeling.

[ 07 ] Inference

Optimized to disappear.

A model is only useful if it answers fast. NEXUS uses speculative decoding, paged attention, and KV-cache quantization to hit 42ms time-to-first-token and stream at 240 t/s.

0
TTFT
0
Throughput
0
Context
0
Quantization

Read the papers.

Six papers, all open. Reproduce, critique, extend.

Research index