February 2026 · Historic First

First LLM inference
on quantum hardware.

Synoptic AI executed the first full transformer inference on gate-based quantum processors — three vendors, two qubit technologies. Classically trained. Quantum executed. Patent pending.

World First · Verified on Hardware

A classically trained transformer generating
real text on real quantum processors

Hello World

Full autoregressive inference — token by token — with every matrix multiplication performed on the QPU. Pretrained weights compressed into quantum-native form via Phase Space and executed as quantum circuits on gate-based processors. Output verified against classical baseline.

144
Quantum Circuits
4–5
Qubits / Circuit
3
Hardware Platforms
2
Qubit Technologies
What this is not
Not a variational circuit trained from scratch on a QPU
Not a simulation
Not a quantum layer bolted onto a classical model
Not an annealer
8–12×
Compression
100%
QPU Accuracy
3
QPU Vendors
20+
Years Research

Verified results on real hardware

Every result below has been validated on gate-based quantum processors. No simulations. No theoretical proposals. Real quantum hardware, real outputs.

Hello World
Gate-Based QPU × 3

LLM Inference on Quantum Hardware

First classically trained transformer to perform full autoregressive inference on gate-based quantum processors. 144 quantum circuits across three hardware platforms, two qubit technologies. Output verified against classical baseline.

100%
Gate-Based QPU

Neural Network Inference on QPU

Classically trained neural network performing inference on a gate-based quantum processor. 10 inputs, 10 correct outputs. Probabilities match classical to 3–4 decimal places.

96%
Gate-Based QPU

Floquet Bivariate Bicycle Code

High-fidelity quantum error correction implementation. Practical error correction on current NISQ hardware toward fault-tolerant quantum architecture.

92%
Gate-Based QPU

SYK Black Hole Microstate Enumeration

Quantum simulation of SYK model black hole microstates. 236/256 microstates with 92% Bekenstein-Hawking accuracy. Extends published research from Brookhaven National Lab.

55%↑
Gate-Based QPU

QAOA Optimization

Proprietary approach achieving 55% improvement over standard QAOA on benchmark optimization problems.

Gate-Based QPU

Kagome t-J Model Simulation

Full-scale condensed matter simulation on Kagome lattice. Physics-validated quantum many-body simulation at utility scale.

1.000
Gate-Based QPU

Porter-Thomas Distribution Validation

Validation of genuine quantum behavior at scale. 1000 shots yielding 1000 unique bitstrings confirms real quantum computation.

Proprietary compression technology

Our patented compression method achieves significant parameter reduction across multiple model scales. The same compressed representation runs on classical hardware today and quantum hardware tomorrow.

8–12×

Parameter reduction with zero translation error

Patent pending. Validated across three model architectures at scales from 124M to 6.7B parameters. The compressed representation is quantum-native — it executes directly on gate-based quantum processors without translation or approximation.

◆ Patent Pending · Method details proprietary
GPT-2 (124M)
11.7×
124M → 10.6M parameters
TinyLlama (1.1B)
8.5×
1.1B → 129M parameters
DeepSeek-Coder (6.7B)
8.2×
6.7B → 826M parameters
→ Classical

Immediate cost reduction

8–12× fewer parameters means proportionally less memory and compute for inference. Applicable to any transformer-based model architecture. Proven across three model scales.

→ Quantum

Native quantum execution

The compressed representation executes directly on gate-based quantum processors without translation or approximation. Validated on quantum hardware with 100% inference accuracy.

Neural architecture innovation

Neural Compression

Proprietary compression techniques achieving significant parameter reduction while retaining intelligence. Scaled and validated on billion-parameter architectures.

8–12× compression · Billion-scale

Mathematical Reasoning Systems

Compact models that solve complex mathematical problems at speeds orders of magnitude faster than frontier systems — competitive accuracy with a fraction of the parameters.

High-speed inference · Minimal parameters

Emergence Detection

Validated frameworks for monitoring and controlling emergence in neural networks — enabling new approaches to AI system behavior and capability assessment.

Novel methodology · Validated

Architectural innovation over brute-force scaling

We develop systems that achieve superior performance through fundamental breakthroughs in how information is processed and represented — not by throwing more compute at the problem.

Two decades of research in intelligence, creativity, cognition, and psychology informs an approach that bridges theoretical insight with practical implementation. Multiple patents pending.

The results speak for themselves. Every claim on this site has been verified on real quantum hardware or against published benchmarks.

Let's talk

Open to partnerships, licensing, collaboration, and consulting arrangements.

michael.hoskins@synopticai.io