Pioneering the era of efficient & secure AI

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 digital natives

Scale without limits

Deploy compact models that eliminate GPU shortages, reduce latency, and accelerate growth.

Learn more

For corporations

Extend your infrastructure

Run advanced AI on existing hardware with compressed models that cut CAPEX and energy use.

Learn more

For data centers

Maximize capacity

Increase throughput and profitability without adding racks — thanks to smaller, faster models.

Learn more

For device manufacturers

Powerful AI on any device

Enable on-device inference with compressed models that fit limited compute and memory.

Learn more

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 CompactifAI removes what the network never used.

Quasar 438B

Parameters: 438B
Output speed: 176 tok/s
Context window: 1M tokens
Input / output per M: $0.60 / $1.80

Artificial Analysis Intelligence Indexv4.1.1 · higher is better

43
Quasar 438B Multiverse Computing · 438B parameters
38
Nemotron 3 Ultra NVIDIA · 550B parameters

13th of 178 models worldwide, and the highest score of any European model — beating a model built with 112 billion more parameters.

Quasar 438B · 13th

#1 worldwide#178

Terminal-Bench 2.1 · agentic coding: 0.0
AA-LCR · long-context reasoning: 0.0

APIcurl api.compactif.ai/v1/chat/completions -d '{"model":"quasar-438b","reasoning":"high", …}'

How CompactifAI 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.

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 Intel Xeon 6 CPUs and Qualcomm 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 SentinelAI either way.

up to 75% less energy, measured by Telefónica

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. SentinelAI 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; Foundry has its own section below.

Deploy compressed AI models anywhere.

Our advanced compression technology reduces LLM size, enabling faster, scalable and cost effective AI on any enterprise system or edge device.

50–80% lower inference costs
Up to 2× faster inference
Close to 100% accuracy retention

Explore CompactifAI

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

See SentinelAI

Optimizing complex problems with Singularity.

Our quantum-inspired optimization platform powers the breakthroughs behind our Compact AI technology.

Portfolio optimization, also inside Excel
Grid, hydrogen and fleet optimization
IBM Quantum partner

Explore Singularity

Foundry · 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.

Foundry eu-west · Donostia 01 / 04 · catalogue early access

Compressed model catalogue4 of 40

Quasar 438Breasoning · agents · coding438B43 AAII
Llama 3.3 70Bcompressed by CompactifAI70Bready
Mistral Largecompressed by CompactifAI123Bready
DeepSeek R1compressed by CompactifAI671Bready

Deploying quasar-438b → on-premready

pulling compressed checkpoint · 21.4 GB
verifying signature · ok
binding SentinelAI policy · redaction, injection screening
warming 4 replicas · Xeon 6

POSThttps://eu-west.internal/v1/chat/completions

Time to first token

0.31 s

Throughput

176 tok/s

Representative panel. Cost and energy figures shown fall inside the 50–80% reduction measured by HP and Telefónica.

The system

One stack. A fraction of the weight.

CompactifAICompress any model, then serve it from one API.

FoundryModels, clusters and cost in one panel.

SentinelAIEvery request inspected, every response audited.

Quasar 438BEurope’s highest-scoring model, built by compressing.

Quantum optimization

Compressed model library

Expertise

Trusted by leading efficiency-driven industries.

Training and inference costs are escalating. With model sizes exploding, efficient AI is no longer optional. Our compression technology makes large-scale AI sustainable.

Green transition

Less compute is less power.

Energy-intensive infrastructures require smarter, more efficient algorithms. Compact AI reduces compute and power usage across critical applications — Telefónica measured up to 75% less energy.

Your sector is not on this list? Wherever compute, energy or regulation is the constraint, the answer is the same shape. Tell us which one is yours. Talk to an engineer

FinanceSovereign LLMs, portfolio and risk optimization

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

Real results

Validated by global leaders.

up to 75% less energy per deployment, running compressed models directly on the network.

The compressed models can be deployed directly on Telefónica’s network, including local facilities, making it possible to reduce energy consumption by up to 75% compared to uncompressed models.”Telefónica · network deployment

50–80% lower cost and energy measured across an extensive internal benchmarking programme.

We conducted extensive technical benchmarking and were very impressed: decrease in time to first token, increase in token throughput, and models that are cheaper to run, with 50–80% cost and energy savings.”HP · benchmark programme

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.Luzia · consumer assistant

independently evaluated by a third party, not by us.

Substantial reductions in power, operational cost and carbon emissions at competitive accuracy.Sopra Steria · independent evaluation

Customers and partners

Company

The biggest quantum software provider in Europe.

Founded in 2019, we recognized quantum technology’s potential to revolutionize AI and optimization solutions. This vision gave rise to Singularity, our advanced platform, created to empower clients. Guided by a customer-centric philosophy, we collaborate to tailor models for specific challenges, aiming to seamlessly integrate user-friendly, disruptive software into the industry landscape.

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.
Singularity · optimization
The same decomposition, applied to portfolio, grid, hydrogen and fleet problems. In production in finance since 2019, and an IBM Quantum partner.
CompactifAI · model compression
And to the weight matrices of the largest language models. 160+ patents stand behind it.

Patents: 160+
Customers: 100+
Series C target, July 2026: $570M
Offices: Europe · US · Canada

Alfonso Rubio-Manzanares, painted portrait

Latest

Launch · 2 Sep 2026
Introducing Quasar 438B, Europe’s leading AI model
Outperforms comparable European models on seven of eight selected Artificial Analysis evaluations.
Partnership · 11 Aug 2026
FIOR Group partnership for secure edge and sovereign AI
Cryptographic agent identity and governance combined with model compression and edge deployment.
Funding · Jul 2026
Series C announced, targeting up to $570M
To expand the library of compressed models and scale commercial deployment worldwide.

Efficient, sovereign AI.

Ten efficiency-driven industries.

The biggest quantum software provider in Europe.

Latest from Multiverse.

From Donostia to the world.

Pioneering the era of efficient & secure AI.

Built for whoever pays the compute bill.

Scale without limits

Extend your infrastructure

Maximize capacity

Powerful AI on any device

Europe’s highest-scoring model. Built by compressing, not by adding.

A 4096×4096 matrix, and the third of it that carries information.

One model. Four places to run it.

Call it. One key, OpenAI-compatible.

Rack it. More tenants per node.

Keep it inside. Data never leaves.

Ship it. It fits in the memory you have.

Compress it. Run it. Host it. Govern it.

Deploy compressed AI models anywhere.

Every request inspected, every response audited.

Optimizing complex problems with Singularity.

Pick a model. Pick where it runs. Watch what it costs.

One stack. A fraction of the weight.

Trusted by leading efficiency-driven industries.

Less compute is less power.

Validated by global leaders.

The biggest quantum software provider in Europe.

Introducing Quasar 438B, Europe’s leading AI model

FIOR Group partnership for secure edge and sovereign AI

Series C announced, targeting up to $570M

All CompactifAI models now run on Intel Xeon 6 and Qualcomm

Up to 75% less energy on Telefónica’s network

HP measures 50–80% lower cost and energy