Train models faster.

At Latent, we develop training algorithms to accelerate AI research. Our goal is to help labs train models faster, run more experiments, and explore new architectures.

01 / THE OPPORTUNITY

Most compute
still goes to training.

63%

pre-training + post-training share
in Q3 2026.

Anthropic + OpenAI combined · SemiAnalysis model

Where the compute goes

Anthropic + OpenAI combined

InferencePost-training / RLPre-training

Q3 2026Post-training / RL 53%Pre-training 10%Inference 37%

Source: SemiAnalysis Tokenomics Model
Anthropic + OpenAI combined. Q4 2026 is an estimate. Labels are rounded.

02 / RESEARCH FOUNDATIONS

The evolution
of optimization.

2014AdamAdaptive learning rates
2017AdamWDecoupled weight decay
2024MuonOrthogonalized momentum
NOWLatentWhat comes next

03 / CONTACT

Let’s speed up
your training.

hello@latent.bot

Train models faster.

Most computestill goes to training.

Where the compute goes

The evolutionof optimization.

Let’s speed upyour training.

Most compute
still goes to training.

The evolution
of optimization.

Let’s speed up
your training.