For digital natives
Scale without limits
No GPU queue to stand in. A compressed model serves more requests on hardware you can actually get, and answers faster while it does it.
FinanceSovereign LLMs, portfolio and risk optimization
Nova 420B is live · OMI 43
Kestrel Energy: 75% less energy
Fold now runs on NPU accelerators
All the intelligence. A fraction of the model.
Quantus compresses large language models with tensor networks, then runs what comes out wherever your data already is: our API, your racks, your devices. The same answers, on a fraction of the compute.
Model size
100%
Accuracy retained
100%
Scroll to compress
up to 95%smaller models
4–12×faster inference
50–80%lower inference cost
Trusted by 90+ organizations
Who runs compressed AI
Built for whoever pays the compute bill.
The same models, four reasons to shrink them: no GPUs to buy, a fleet to fill, racks to keep full or a device with almost no memory.
For corporations
Extend your infrastructure
The servers you already bought are the deployment. Frontier-class models fit on them once they stop carrying weights nobody uses.
For data centers
Maximize capacity
Same racks, more tenants. Throughput per node rises because each model takes a fraction of the memory, and of the power.
For device manufacturers
Powerful AI on any device
Inference where the data is made: a controller, a vehicle, a handset. No round trip, no per-token bill, nothing leaving the device.
Released 2 September 2026
Europe’s highest-scoring model. Built by compressing, not by adding.
A reasoning model for enterprise agents and coding. It beats models with more parameters because Fold removes what the network never used.
Nova 420B
- Parameters
- 420B
- Output speed
- 176 tok/s
- Context window
- 1M tokens
- Input / output per M
- $0.60 / $1.80
Open Model Indexv4.1.1 · higher is better
- 43 Nova 420B
- 38 Helion 2 Ultra
13th of 178 models worldwide, and the highest score of any European model — beating a model built with 112 billion more parameters.
Nova 420B · 13th
#1 worldwide#178
- AgentTerm 2.1 · agentic coding
- 0.0
- LCR-9 · long-context reasoning
- 0.0
API
curl api.compactif.ai/v1/chat/completions -d '{"model":"nova-420b","reasoning":"high", …}'How Fold works
A 4096×4096 matrix, and the third of it that carries information.
Tensor networks are the framework physicists use for quantum many-body systems. Pointed at a weight matrix, they show which correlations carry signal and which are noise. What survives is decomposed into a chain of small tensors: the same model, a fraction of the weight.
A dense 4096×4096 weight matrix is profiled layer by layer; 66% of its correlations are found to be spurious and discarded; the remaining third is decomposed into a chain of five small tensors with bond dimension 64, then healed by a short retraining pass. The result has 95% fewer parameters and retains 97.4% of the accuracy, running 4 to 12 times faster at 50 to 80% lower cost.
One model. Four places to run it.
The same compressed weights move from our API to your data center to a device without changing a line of your integration. Choose by where the data lives, not by where the GPUs are.
01Cloud API
01 · Cloud API
Call it. One key, OpenAI-compatible.
The compressed model lives at an endpoint served from Europe. Structured outputs, tool calling and always-on reasoning with effort controls.
TTFT 0.31 s · 176 tok/s
02 · Data center
Rack it. More tenants per node.
The same weights, exported for your servers. Runs on commodity x86 CPUs and NPU accelerators as well as GPUs, so capacity grows without new hardware.
4–12× faster inference per node
03 · Edge and on-prem
Keep it inside. Data never leaves.
Sovereign deployments for banks, utilities, defense and the public sector. Air-gapped if you need it, governed by Warden either way.
up to 75% less energy, measured by Kestrel Energy
04 · On device
Ship it. It fits in the memory you have.
Small enough for an industrial controller, a vehicle or a phone. Inference happens where the data is produced, with no round trip to a cloud.
0 bytes to the cloud
One model. Anywhere you need it.One compressed checkpoint, exported for GPU, CPU and NPU runtimes. The same OpenAI-compatible endpoints in your rack as in our cloud. Warden in front of all of it.
The AI Suite
Compress it. Run it. Host it. Govern it.
Four products built on one idea: AI that fits the hardware, the budget and the jurisdiction you already have. Three are here; Forge has its own section below.
Compress it once. Run it anywhere.
Tensor networks find the correlations a weight matrix never uses and throw them away. What is left is a model up to 95% smaller that answers like the one you started with — and an API that serves it.
- 50–80% lower inference costs
- Up to 2× faster inference
- Close to 100% accuracy retention
Every request inspected, every response audited.
One governed pipeline between every application and every model. Aligned with the EU AI Act, GDPR and SOC 2, adaptable to ISO, HIPAA and DORA.
- Redacts PII and secrets before they leave
- Screens prompt injection and policy breaches
- Observes spend and risk per call
- Audits every request and response
The solver the compressor is built on.
Choosing what to discard is itself an optimization problem. The annealing engine that solves it also solves the ones our customers bring: portfolios, grids, fleets, schedules.
- Portfolio and risk, in the tools finance already uses
- Grid, hydrogen and fleet optimization
- Runs on CPUs today, quantum hardware when it lands
Forge · early access
Pick a model. Pick where it runs. Watch what it costs.
A unified platform for sovereign AI infrastructure: explore models, deploy GPU clusters, monitor costs and workloads, and access serverless AI services — all within a clean, intuitive control panel. Try the four steps.
Forge
eu-west · Lisbon
01 / 04 · catalogue
early access
Compressed model catalogue4 of 40
- Nova 420Breasoning · agents · coding420B43 AAII
- Orion 3.3 70Bcompressed by Fold70Bready
- Calder Largecompressed by Fold123Bready
- Tessera R1compressed by Fold671Bready
Where it runsdata residency: EU
- Your data centerx86, GPU or NPUon-premselected
- Cloud APIserved from Europemanagedavailable
- Air-gappedno route to the internetsovereignavailable
- On deviceindustrial controller, vehicle, handsetedgeavailable
Deploying nova-420b → on-premready
- pulling compressed checkpoint · 21.4 GB
- verifying signature · ok
- binding Warden policy · redaction, injection screening
- warming 4 replicas · x86 nodes
POST
https://eu-west.internal/v1/chat/completionsLast 30 daysvs. the uncompressed model
uncompressed baselinecompressed, after cut-overcost index · baseline = 100
Representative panel, invented figures — like everything else on this page.
The system
One stack. A fraction of the weight.
Expertise
Built for the industries that count watts.
Model sizes grow faster than the budgets and the grid connections that have to carry them. Wherever compute, energy or regulation is the binding constraint, the same answer applies: make the model smaller before you make the data center bigger.
Green transition
Less compute is less power.
A model that fits in a tenth of the memory draws a fraction of the power, every request, for as long as it runs. Kestrel Energy measured up to 75% less energy after the switch.
EnergyGrid, storage and demand optimization
ManufacturingEdge inference on the line
Health & Life SciencesPrivate models on hospital hardware
EngineeringDigital twins and simulation
AerospaceFlight-operations optimization
CybersecurityGoverned, air-gapped inference
DefenseSecure-by-design AI chips
ChemistryMolecular and process modeling
HydrogenStorage and transport optimization
Results
Validated where it actually runs.
up to 75% less energy
per deployment, running compressed models on hardware already in the substations.
The compressed models run on the hardware already sitting in our substations. Energy per inference fell by up to 75% against the models they replaced.”Kestrel Energy · grid operations
50–80% lower cost and energy
measured across four months of internal benchmarking.
We benchmarked it for four months: time to first token down, throughput up, and 50–80% less cost and energy for every request we serve.”Halden · platform engineering
over 50% smaller footprint
in a consumer support assistant, at the same response quality.
Model footprint cut by over 50% in a consumer support assistant, with lower latency and cost at the same response quality.Luma · consumer assistant
independently evaluated
by a third party, not by us.
Substantial reductions in power, operational cost and carbon emissions at competitive accuracy.Nordaudit · independent evaluation
Customers and partners all invented
Kestrel Energy
Halden
Terrabank
Luma
Aether Cloud
Nordaudit
Telaris
Silvachem
Arden Motors
Meridian Health
Navis Marine
Caldera Semi
Ponte Infra
Heliogrid
Ferrova Rail
Deltawerk
Quinta Aero
Ossian Pharma
Boreal Defence
Civitas
Nimbus Logistics
Aquifer H2
Astra Grid
Consortium Nova
Company
Mathematics from physics, pointed at the weight of models.
Quantus builds on one piece of mathematics: the tensor network, invented so physicists could describe systems too large to write down. It turns out a weight matrix is such a system, and most of what it stores is correlation nobody uses. Everything here — the solver, the compressor, the models, the platform — is that one idea, applied a layer at a time.
-
Quantum many-body physics
Tensor networks were invented to describe systems with more states than there are atoms in the universe — by keeping only the correlations that matter.
-
Anneal · optimization
The same decomposition, applied to portfolio, grid, hydrogen and fleet problems — where the search space is far larger than the time anyone has to search it.
-
Fold · model compression
And to the weight matrices of the largest language models: keep the correlations that carry signal, drop the two thirds that do not, retrain briefly to heal the seams.
- Patents
- 140+
- Customers
- 90+
- Raised to date
- $310M
- Offices
- Lisbon · Zurich · Montreal
Latest
Introducing Nova 420B, our highest-scoring model
Outperforms comparable open models on seven of the eight evaluations in the Open Model Index.
Aether Cloud partnership for sovereign edge deployment
Cryptographic agent identity and governance, combined with model compression and edge rollout.
Series C closed at $180M
To expand the library of compressed checkpoints and scale commercial deployment.
Every Fold model now runs on commodity CPUs and NPUs
As well as GPUs, so capacity grows without buying new hardware.
Up to 75% less energy across the substation estate
Compressed models deployed on hardware that was already installed on site.
Halden measures 50–80% lower cost and energy
After four months of benchmarking: faster time to first token and higher throughput.
The part that is not fiction
Designed and built by agents. Twenty-four rounds of it.
Quantus is invented — the name, the mark, the products, the customers and every figure quoted above. The work is not. One hand-written HTML file, fifteen illustrations generated from code instead of drawn, a mark whose geometry feeds three different renderers, and a measuring tool written for every claim made here.
- Rounds of human feedback
- 24
- The page
- 1 file
- Illustrations, generated
- 15
- Measurement tools written
- 87
01
Drawn by code, not by hand.
The fifteen isometric scenes are Python: solids placed in world units, painted back to front, lit by one light. The mark is generated too, in three forms — the logo, the nine-band version printed on the top face of every drawing, and a cell grid that the footer's 512 cubes fly into. Change the geometry once and all three follow.
02
Nothing is done until it is measured.
Every round ends in real browsers: Chrome and WebKit, at 1440 and at 390, console clean, no response ≥400, no horizontal overflow. Performance is numbers, not adjectives. The round that went after scroll jank moved main-thread work from 40.1 s to 26.4 s, and frames over 20 ms from 2.5% to 0.9%.
03
Most of it is deciding what not to keep.
Seventy-three decisions sit on a list headed do not go back: cubes behind the buttons, a Rubik's cube in the footer, every elastic bounce, stock photography in the news. Three obvious optimizations were measured, found worse, and reverted.