Powering the SI era

Looped transformers.
Growing memory.

Spotlight is a looped transformer with growing memory — depth that deepens at test time, and memory that compounds across sessions.

Try the Playground Read the Paper
CA · Robinhood
ModelLoop-8 Tokens/s4.2M MemoryGrowing Uptime99.98% Checkpoint9/9/26 BlogSpotlight Weekly · Sep 30

The Wall

Forward pass: one and done

Billions in compute. Zero memory.

Memory a static model retains after its context closes

0

tokens remembered across sessions by a frozen forward pass

Dense Transformer

Dense · Frozen

No memory

Mixture of Experts

Sparse · Static

No recurrence

Hybrid SSM

Fixed · Bounded

Bounded memory

Spotlight

Looped · Growing

Adaptive

Test-time recurrence

Every architecture above learns once, then forgets forever. Spotlight was built to keep going after deployment.

See the architecture

Source: Spotlight Research · recurrence & memory benchmarks · 2026

Why Spotlight

Spotlight is a looped transformer with growing memory. It reuses its depth at test time and writes what it learns into persistent memory, so capability compounds instead of resetting.

Architecture

CoreLooped transformer DepthAdaptive · test time MemoryPersistent · growing StatusLearning

Performance

Throughput4.2M tok/s Loop depth8 · adaptive ContextUnbounded Latency6 ms

Roadmap

TrainingContinual WeightsOpen · 2026 AccessPublic beta AlignmentLoop interpretability

The Model

What if intelligence never forgot?

Looped by design

Spotlight re-runs its core at test time, spending compute on the tokens that matter. Depth adapts, reasoning deepens, cost stays flat.

Read the Paper

Early access (coming soon)

The Playground opens soon. Watch Spotlight reason through loops and grow its memory in real time — then deploy it behind an API.

Join the waitlist

Memory that grows

Every session.
Every task.
Nothing forgotten.

Read the Paper

The Ecosystem

Tools. Entry points to Spotlight.

Platform

Run Spotlight on your stack.

247 active loops

View API

Community

Get involved.

X / Twitter

Notes from the lab, as they ship

Follow on X

Latest from Spotlight

Stay ahead of the intelligence shift.

Blogs

Sep 30Weekly Update

Spotlight Weekly: The Memory Keeps Growing

Three weeks of continual learning: memory grew 14%, retention held at 96%, and Loop-8 matched a dense 70B on reasoning evals.

Sep 16Weekly Update

Spotlight Weekly: Loops Beat Depth

Our looped core matches a 70B dense model at a fraction of the parameters — and keeps improving at test time while static models stay frozen.

Aug 26Research

Out of Conversations: Why Recurrence Beats Depth

Stacking layers is a one-way street. Looping a small core turns compute into thinking — and why that changes what scaling means.
View all posts →

The Loop Podcast

The Loop #24 Loops, Memory, and the SI Timeline

Latest episode

#23 — In Conversation — Memory That Grows at Test Time

2 months ago

Watch on YouTube →
View all episodes →

24 episodes on The Loop