Introducing Fusion Training

Only five companies can afford to build frontier AI. We think that's a bug.

The problem

Intelligence is gated by capital, not ideas.

Training a frontier model costs $100 million to $1 billion. It demands purpose-built data centers — thousands of specialized chips, custom networking, and megawatts of continuous power.

The open-source community has the talent and the will, but not the hardware. Universities, startups, developing nations, and independent researchers are locked out of the most important technology of the century. The barrier was never ideas. It was money.

$100M–$1B per model Thousands of GPUs Megawatts of power 5–6 organizations

Everyone is optimizing the same trade. Fusion Training abandons it.

A new category

It doesn't fit the existing buckets.

Every other method either rebuilds the model from scratch or leaves its capacity untouched. Fusion Training grows a trained model — adding real new neurons and connections — while preserving everything it already knows.

Pre-training
Builds a model from scratch.
$100M+ · Months
Fine-tuning
Redirects what the model already knows. Cheap, but adds no new capability.
Cheap · No new capacity
Continued pre-training
Keeps training on more data. Same model, same size, same architecture.
Expensive · No new capacity
Fusion Training
Grows the model. Adds real new neurons, trains only the new parts, preserves everything else.
$3,000 desktop · 20 min
How it works

Diagnose. Expand. Train. Optimize. Repeat.

Think of a model like a brain. Today, a smarter brain has to be grown from scratch in an expensive factory. Fusion Training grows the brain you already have — surgically, and on a desktop.

01
Diagnose

Scan the model to see which parts are working hard and which have spare capacity. The result is a map of exactly where it needs more intelligence.

2 min
02
Expand

Surgically add new neurons where capacity is needed. A function-preserving technique guarantees the model behaves exactly as before — nothing lost, nothing broken. It's just bigger now, with room to learn.

12 sec
03
Train

Train only the new neurons. The original model stays frozen. Because just a few layers learn at a time, you need a fraction of the compute traditional training demands.

20 min
04
Optimize

Scan the larger model for dead weight — neurons left over from the original training that aren't pulling their weight — and remove them. The model gets smaller and faster with zero quality loss. We call it inverse fusion.

Contraction
05
Repeat

Each cycle makes the model smarter. Run it nightly, weekly, or continuously. The model evolves over time instead of being rebuilt from scratch every 6–12 months.

Nightly

A $4,000 desktop matched a $25,000 GPU. To within 0.04%.

What we proved

Validated across 26 sessions.

Every experiment ran on an NVIDIA DGX Spark — a desktop computer that draws 65 watts and sits on a desk — compared against an NVIDIA A100, the $25,000 professional GPU that is the industry standard. Same starting model, same data, same number of steps. The only difference was the method.

99.96%
of the A100's quality, matched
faster to train
the hardware cost
the power draw
Method
Perplexity ↓
Parameters
Starting model
29,183
1.55B
A100 — traditional training
9.76
1.55B
Fusion — resource-efficient mode
11.54
2.6B
Fusion — quality-maximized modeBeats A100
9.24
2.6B

Lower perplexity is better. Fusion Training didn't just match quality — given equivalent compute, it built a larger, more capable model (1.55B → 2.6B parameters) that beat the A100 outright, because the expanded architecture had more room to learn.

Growth cycles compound

We ran 6 consecutive growth cycles. Quality improved every single time. The model gets smarter with every iteration.

Contraction works

After growing the model, we identified and removed dead weight — reducing size by 10% with zero quality loss. Smarter and more efficient.

Architecture-agnostic

Validated on a non-standard hybrid architecture combining Transformer and Mamba layers. It worked on both. Standard models like Llama, Qwen, and Mistral are simpler.

Scales without limit

Fusion trains a few layers at a time, using ~24 GB of memory regardless of total model size. A 1B model and a 1T model need the same memory — only the time changes.

Why it matters

The industry's core assumption was wrong.

Capital

Frontier training currently costs $100M–$5B. Fusion reduces it to the cost of a desktop and electricity. Any organization can grow AI models.

Energy

Comparable results at 1/600th to 1/4000th the energy. A 65-watt desktop replaces a rack of GPUs drawing kilowatts — the compliant path as regulators scrutinize AI's power appetite.

Continuous improvement

Today's models are frozen snapshots, replaced every 6–12 months at enormous cost. Fusion models improve nightly. No more launches — the model simply gets better every day.

Distributed training

No GPU-to-GPU communication required. A thousand desktops on ordinary internet can collectively train a model no single data center could afford to build from scratch.

Fusion Training

Intelligence, no longer gated by capital.