Layer compression
Single dense layer, tensor-train
200,000×
Denseχ=64χ=16χ=4χ=2χ=1
Parameters, log scale. Schematic for a 25088×4096 layer.
arXiv:1509.06569Quantum-native AI
Generative models
built on quantum math
Quantaria builds multimodal models whose weights are tensor networks from the first training step. Text and images come from one shared backbone, in the same format physicists use for quantum states.
Hilbert-1 Nano and Base are live on the console Open the console See the models- MPS/MPO weights from step zero
- Bond dimension χ set per layer
- Never compressed after the fact
- Next-token loss for text
- Diffusion loss for image patches
- One set of tensor cores
- Bond dimensions adapt during training
- Schmidt-rank truncation per bond
- Canonical-form gauging for stability
- MPS cores compile to staircase circuits
- Disentangler layers on simulators first
- Hybrid layers tested on IBM Heron in published work
|ψ(x)|² · interference of 4 sources
Rendered live, 1-bit ordered dither
Overview
Quantaria builds multimodal generative models whose weights are tensor networks from the first training step. Text and images come from one shared backbone.
Every weight sits in the same format physicists use for quantum states, so bond dimension sets capacity the way entanglement does, and layers have a direct path from GPU to QPU.
Hilbert-1 · training pipelinePipeline example
- 01Factor weights
- 02Train rank-adaptive
- 03Truncate Schmidt rank
- 04Compile to circuit
-
Tensor cores
MPS/MPO weights trained natively.
-
Shared backbone
Text and images from one set of cores.
-
Rank control
Bond dimension adapts as it trains.
-
QPU path
Cores map onto quantum circuits.
Capabilities
Six parts, one tensor-native stack.
Each part is proven in published research. Quantaria puts them together in one model that generates text and images.
Layer inventory
| Layer | Format | χ | Status |
|---|---|---|---|
| Embedding | MPS | 32 | Trained |
| Attention (TPA) | Tensor product | 64 | Trained |
| Experts | MPO shared core | 48 | Trained |
| Feed-forward | Tensor train | 64 | Trained |
| Diffusion head | Tensor train | 32 | Trained |
| Hybrid layer | Circuit + MPO | 4 | Simulator |
from bond import TTLinear layer = TTLinear(4096, 4096, cores=4, chi=64) layer.params() # 532,480 vs 16,777,216 dense
Illustrative layer inventory for Hilbert-1 Base. Bond is Quantaria’s planned open-source library.
Numbers
What tensor networks
have already done
Published results on the methods Quantaria combines. Quantaria's own benchmarks replace these as each model is evaluated.
Memory
LLaMA-7B
Memory reduction93%
Accuracy cost2–3%
Formattensor network
KV cache
Tensor Product Attention
~10×
MERA layers
Image classifiers
14,000× fewer params
<1% accuracy loss
arXiv:1711.03357Hybrid quantum layer
SmolLM2 attention layer
~3,000×
Classical
Hybrid
+0.26% perplexity
arXiv:2410.17397Experts
vs Switch Transformer
27× fewer params
11427
arXiv:2203.01104
Training memory
vs GaLore
9×
2–3× faster epochs
arXiv:2405.14377GPT-2 in TT form
Near-baseline perplexity
~40% fewer params
arXiv:2306.02697Models
Five model lines, named for the physics they run on
Hilbert-1 Nano and Base are live on the Quantaria console. The other lines are in research or planned.
Hilbert
001Schmidt
002Dirac
003Born
004Bond
005Specs
Hilbert-1, three sizes
Nano and Base are live on the Quantaria console. Large follows Base results.
|
Compare model sizes
|
Nano ~100M The smallest Hilbert-1. Text and images from one tensor backbone. Open console | Base ~1B Scaled up with MPO experts. Live on the console. Open console | Large TBD Scale set after Base results. |
|---|---|---|---|
| Architecture | |||
| Parameters | ~100M | ~1B | Set after Base |
| Text context | 2K tokens | 8K tokens | TBD |
| Image output | 32×32 | 256×256 | TBD |
| Stack | |||
| MPS/MPO weights | |||
| Tensor Product Attention | |||
| Transfusion objective | |||
| Rank-adaptive χ | |||
| MPO mixture-of-experts | – | ||
| Hybrid quantum layers | – | – | Research |
| Evaluation | |||
| Dense baseline, same size | |||
| Training data | TinyStories + CIFAR-scale | TBD | TBD |
| Benchmark report | Coming soon | Coming soon | After release |
| Status | |||
| Availability | ConsoleLive | ConsoleLive | After BasePlanned |
| Where to use it | console.quantaria.xyzOpen now | console.quantaria.xyzOpen now | –Planned |
Quantum path
From GPU to QPU,
without changing the format
Tensor networks are the native language of quantum circuits, so layers can move to quantum hardware as it matures.
Same weights, three backends
Tensor cores
MPS/MPO layers train and run on GPUs today.
Fused contractions
Cores contracted in fused kernels.
Rank control
χ adapts per bond during training.
KV cache
~10× smaller with tensor product attention.
Shared backbone
Text and images from one set of cores.
Baselines
Every model vs a dense model of the same size.
Staircase circuits
MPS with χ = 2k → gates on k+1 qubits.
Disentanglers
MERA-style circuits absorb correlations from the MPO.
Bond-1 remainder
~3,000× fewer classical params in one attention layer (arXiv:2410.17397).
Unitary adapters
Cayley-parameterized, exactly unitary while training.
Entanglement budget
χ bounds entropy at log2χ bits per cut.
Noise models
Layers tested against simulated hardware noise.
IBM Heron
Part of a hybrid layer has run on superconducting hardware in published work.
Hybrid layers
Quantum circuit + MPO + quantum circuit.
Cloud QPUs
Next step after simulators.
Only where it helps
Quantum layers ship only where real hardware supports them.
Born sampling
Sampling |ψ(x)|² like a measurement.
Area-law ranks
Rank set by measured entanglement entropy.
Quantaria models train and run on GPUs today. Quantum layers run on circuit simulators and small hardware experiments.
Roadmap
Prove it small,
then scale it
Each step publishes its results, including the ones that don't work.
Hilbert-1 Nano and Base
Live on the Quantaria console. Benchmarks against dense transformers of the same size follow: loss curves, perplexity, FID and memory use.
Bond, Schmidt and Hilbert-1 Large
Open-source Bond. Train Schmidt for on-device long context and scale Hilbert to Large.
Dirac hybrid layers
Swap selected attention layers for disentangler circuits on simulators, then cloud quantum hardware.
Quantum-native training
Area-law rank allocation, DMRG-style sweeps alongside gradient descent, Born-machine heads.
Matched baselines
Every model vs a dense transformer of the same size, same data.
Cited numbers
Every figure on this site links to its paper.
Hardware-honest
We say plainly when a result came from a simulator.
FAQ
Frequently asked questions
Questions about the research, the models or working with us?
Contact usA research company building multimodal generative models whose weights are tensor networks (MPS/MPO) from the first training step. One backbone generates both text and images.
Not today. Quantaria models train and run on GPUs. Because the weights are in the same format as quantum states, individual layers can move to quantum hardware as it matures. Hybrid layers are tested on simulators first.
No. They come from published research on the methods Quantaria combines, and each links to its paper. Quantaria's own benchmarks will be published as each model is evaluated.
Yes. Hilbert-1 Nano (~100M parameters) and Hilbert-1 Base (~1B parameters) are live on the Quantaria console at console.quantaria.xyz. Both generate text and images from one tensor-network backbone. Hilbert-1 Large follows Base results.
Yes. Bond, the PyTorch library behind every Quantaria model, is planned as an open-source release.
Email hello@quantaria.xyz.
Hilbert-1 is live
Smaller weights.
Quantum path.
Tensor-network-native models for text and images, built in the format quantum hardware speaks.
- MPS / MPO
- TPA
- MERA
- DMRG