quarterbit 4.0.3

AXIOM - World's most memory-efficient drift-free optimizer for PyTorch

QuarterBit AXIOM

World's Most Memory-Efficient Drift-Free Optimizer

90%+ memory savings. Zero precision loss. Train larger models.

The Problem

Training large AI models requires massive GPU memory for optimizer states. Adam stores 8 bytes per parameter just for momentum and variance - that's 8GB for a 1B parameter model.

Additionally, long training runs suffer from floating-point drift - tiny gradient updates accumulate rounding errors over millions of steps, causing:

• Stalled convergence in late training
• Numerical instability
• Suboptimal final models

The Solution

QuarterBit AXIOM solves both problems:

Metric	Adam	AXIOM	Improvement
Memory per param	8.0 bytes	0.76 bytes	90%+ savings
Precision drift	Accumulates errors	Drift-free	Eliminated
Max model (16GB GPU)	~1.0B params	~2.7B params	2.7x larger

Installation

pip install quarterbit

Supported GPUs:

• NVIDIA T4, V100, A100, L4, L40
• NVIDIA RTX 30 series (3060-3090)
• NVIDIA RTX 40 series (4060-4090)
• NVIDIA H100, H200

Quick Start

from quarterbit import Axiom

# Drop-in replacement for Adam
optimizer = Axiom(model.parameters(), lr=1e-3)

# Optional: set training schedule for adaptive LR
optimizer.set_schedule(total_steps=10000, warmup_steps=1000)

# Train as usual
for batch in dataloader:
    loss = model(batch)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

Why AXIOM?

1. Massive Memory Savings

Train 2.7x larger models on the same GPU. AXIOM uses proprietary compression to reduce optimizer state from 8 bytes to under 1 byte per parameter.

2. Drift-Free Training

Proprietary precision algorithms eliminate floating-point accumulation errors. Your model trains with perfect numerical stability from step 1 to step 1,000,000+.

3. Cloud Cost Reduction

Less memory = fewer GPUs = lower costs. Typical savings of 30-50% on cloud training bills.

4. Drop-In Replacement

Same API as PyTorch Adam. Change one line of code.

Benchmarks

Validated on GPT-2 (124M parameters):

Optimizer	Final Loss	Memory	Status
AdamW	3.12	1,024 MB	Baseline
8-bit Adam	3.14	640 MB	-37%
Adafactor	3.18	512 MB	-50%
AXIOM	3.11	95 MB	-91%

See our Kaggle benchmark for full results including OOM stress tests and cost analysis.

Requirements

• Python 3.8+
• PyTorch 1.8+
• NVIDIA GPU with CUDA support
• Linux or Windows

Pricing

Tier	Price	Use Case
Free	$0	Personal, research, evaluation
Pro	$299/mo	Commercial use, up to 10 GPUs
Team	$2,499/mo	Up to 100 GPUs, priority support
Enterprise	Custom	Unlimited, custom SLA

See quarterbit.dev/pricing for details.

License

Proprietary - see LICENSE for details. Free tier available.

Links

• Website: quarterbit.dev
• Documentation: quarterbit.dev/docs
• Email: [e-mail]

---

Copyright (c) 2026 Clouthier Simulation Labs. All rights reserved.