---
title: "Quasar 438B Is API-Only, and That Is Not Sovereignty"
slug: "multiverse-quasar-438b-european-sovereign-ai-api-only"
author: "Mikel Amigot"
date: "2026-09-10 17:00:00"
category: "Premium"
topics: "European AI, digital sovereignty, Quasar 438B, Multiverse Computing, open-weight models, EU AI Act, model-agnostic AI"
summary: "Multiverse Computing's Quasar 438B scored 43 on Intelligence Index v4.1.1 at launch on 2 September 2026, the top European result. It is also proprietary, API-only, and compressed from Z.ai's open-weights GLM-5.2."
banner: ""
thumbnail: ""
linkedin: |
  Spain's Multiverse Computing launched Quasar 438B on 2 September 2026, and at launch it was the highest-scoring European model on the Artificial Analysis Intelligence Index — 43 on v4.1.1, against 38 for NVIDIA Nemotron 3 Ultra and 30 for Mistral Medium 3.5.

  That is a real result, and European buyers should be pleased about it. It is not digital sovereignty, and the two keep getting conflated.

  Two facts from Multiverse's own material decide it. The weights are proprietary, and the model is served only through the CompactifAI API — it does not appear among the models Multiverse offers for private deployment. And Quasar is a compressed build of GLM-5.2, the open-weights model from Z.ai, with the expert count pruned from 265 to 148 per layer.

  So a European ministry running Quasar sends its prompts to someone else's inference endpoint, cannot inspect the weights, and is downstream of a base model trained outside Europe.

  → Sovereignty is decided by where inference runs, not by the vendor's postcode
  → It is decided by who holds the weights and the data at rest
  → It is decided by whether you can switch models without rewriting the platform
  → A benchmark score decides none of those three

  With ibl.ai you own all the code and the data — self-hosted inside your own perimeter, model-agnostic across any LLM, usage-based with no per-seat pricing, deployable anywhere from your own cloud to a fully air-gapped network.

  #iblai #AgenticAI #EnterpriseAI #DigitalSovereignty #EuropeanAI #EUAIAct
---

## The Short Answer

**Multiverse Computing launched Quasar 438B on 2 September 2026, and at 43 on Artificial Analysis Intelligence Index v4.1.1 at launch it was the highest-scoring European model. It is also proprietary, API-only, and compressed from Z.ai's open-weights GLM-5.2 — so it is not sovereignty. Sovereignty is decided by where inference runs and who holds the weights. With ibl.ai you own all the code and the data.**

A good European model and a sovereign European deployment are different purchases, and the launch coverage merged them.

## What is Quasar 438B, and what did Multiverse Computing actually ship?

A 438-billion-parameter reasoning model from Multiverse Computing, headquartered in Donostia-San Sebastián, Spain, [announced on 2 September 2026](https://www.globenewswire.com/news-release/2026/09/02/3355465/0/en/multiverse-computing-launches-quasar-438b-the-highest-scoring-european-model-on-artificial-analysis-intelligence-index.html).

The headline claim is the Artificial Analysis Intelligence Index, where Multiverse reports a score of **43** on v4.1.1 — ahead of **NVIDIA Nemotron 3 Ultra at 38** and **Mistral Medium 3.5 at 30**, and the highest among the European models in that comparison.

The supporting numbers are specific enough to check. Terminal-Bench v2.1 of **69.3**, an AA-LCR long-context reasoning score of **75.0**, and a 500-token response including thinking time in **15.3 seconds**.

Artificial Analysis independently lists the served model at roughly [**180 output tokens per second**, a 1M-token context window, and $0.60 per million input tokens against $1.80 per million output tokens](https://artificialanalysis.ai/models/quasar-438b/providers).

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Property</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Published value</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Launch date</td><td style="padding:0.75rem;">2 September 2026</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Parameters</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">438B</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">AA Intelligence Index (v4.1.1)</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">43</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Terminal-Bench v2.1</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">69.3</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Context window</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">1M tokens</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Output speed</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">~180 tokens/sec</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Price (in / out)</td><td style="padding:0.75rem; font-variant-numeric:tabular-nums;">$0.60 / $1.80 per 1M tokens</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Languages</td><td style="padding:0.75rem;">English, Spanish</td></tr>
    <tr style="border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;">Base model</td><td style="padding:0.75rem;">GLM-5.2 (Z.ai, open weights)</td></tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;"><td style="padding:0.75rem;"><strong>Weights</strong></td><td style="padding:0.75rem;"><strong>Proprietary, not published</strong></td></tr>
    <tr style="background:#f0f9ff;"><td style="padding:0.75rem;"><strong>Access</strong></td><td style="padding:0.75rem;"><strong>CompactifAI API only</strong></td></tr>
  </tbody>
</table>

## Is Quasar 438B open-weights, and where do its weights come from?

No, and this is the part that most of the launch coverage left out.

Multiverse's own technical write-up states it plainly: [Quasar 438B is a compressed model built from GLM-5.2, the open-weights model from Z.ai](https://multiversecomputing.com/papers/inside-quasar-438b-by-multiverse-computing).

The compression is the company's actual product. A quantum-inspired pruning pass cuts the expert layers from **265 to 148 experts per layer**, followed by a healing pass and quantization-aware compression in FP8 and NVFP4.

That is legitimate engineering, and the resulting model is genuinely faster and cheaper to serve than the base. It is also not a European base model. The trained weights Quasar derives from came from a Chinese lab, and the Quasar weights themselves are not published.

[Quasar is announced only through the CompactifAI API](https://multiversecomputing.com/resources/introducing-quasar-438b-europe-s-leading-ai-model) and does not appear among the models Multiverse offers for private deployment.

In fairness to Multiverse, CompactifAI offers to create a custom "Slim" version of an open-source model a customer already uses. That is a different product from Quasar, and it is the one a sovereignty-minded buyer should be asking about.

## Does running a European model make a European deployment sovereign?

No. A model's nationality is a property of the vendor, and sovereignty is a property of the deployment.

If your ministry sends prompts to `dashboard.compactif.ai`, then your inference runs on infrastructure you do not control, your data leaves your perimeter in transit, and your continued access depends on a commercial relationship with one supplier — the sole provider serving this model.

None of that changes because the supplier is in Spain rather than California. It is a shorter jurisdictional distance, which is worth something, and it is not the same thing as control.

The three questions that actually decide it are: where does inference execute, who holds the weights and the data at rest, and can you replace the model without rewriting the system around it.

This is the same test we applied to [digital sovereignty in government AI](/blog/government-ai-digital-sovereignty-model-agnostic), and Quasar passes none of the three today.

Regulatory pressure makes the distinction concrete rather than academic.

The EU AI Act [entered into force on 1 August 2024, with obligations for general-purpose AI models applying from 2 August 2025](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) — obligations a public body has to evidence for the system it operates, not for the vendor's headquarters.

## What should a European ministry or bank do about Quasar 438B today?

Use it, if it is the best model for the task, and stop treating that decision as a sovereignty decision.

Quasar is priced at a level where evaluating it costs almost nothing, and a 1M-token context with a 69.3 Terminal-Bench score is a serious agentic-coding proposition. Route the workloads that suit it there.

Then keep the sovereignty question where it belongs: in the platform layer that sits above the model. That layer is what holds your data, your permissions, your audit trail and your integrations, and it is the piece you cannot swap on a quarter's notice.

The practical shape is a platform you run yourself that can call Quasar over the API today, run [Mistral's open-weight models](/blog/mistral-ai-for-education-european-open-source-excellence) or Gemma or Llama on your own hardware tomorrow, and move a workload between the two without a rewrite.

A model is a component with a shelf life of months. The platform around it is a decade-long commitment, which is why the ownership question belongs there.

## How does ibl.ai make a European AI deployment sovereign?

By separating the two decisions the Quasar coverage conflated: which model you use, and who controls the system running it.

With ibl.ai you own all the code and the data.

The platform is delivered as full source code under a perpetual license and runs inside your own perimeter — your cloud, your VPC, on-premise, or a fully air-gapped network with no outbound connectivity.

It is model-agnostic across any LLM, so Quasar over the CompactifAI API and a self-hosted open-weight model can run side by side and swap without touching application code.

Billing is usage-based with no per-seat pricing, so cost tracks what your organization actually processes rather than how many civil servants have accounts.

For a European buyer, the practical consequence is that the residency question has a factual answer: the inference endpoint is one you chose, and the data at rest sits in storage you administer.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY.

*Related reading: [digital sovereignty: why agencies need model-agnostic AI](/blog/government-ai-digital-sovereignty-model-agnostic) — the same three-question test, applied to public-sector procurement.*

*Sources: launch date, Intelligence Index score of 43, Terminal-Bench 69.3 and the Mistral/Nemotron comparisons from [Multiverse Computing's release](https://www.globenewswire.com/news-release/2026/09/02/3355465/0/en/multiverse-computing-launches-quasar-438b-the-highest-scoring-european-model-on-artificial-analysis-intelligence-index.html) and [its launch page](https://multiversecomputing.com/resources/introducing-quasar-438b-europe-s-leading-ai-model); the GLM-5.2 base model and the 265→148 expert pruning from [Multiverse's technical write-up](https://multiversecomputing.com/papers/inside-quasar-438b-by-multiverse-computing); output speed, context window and pricing from [Artificial Analysis](https://artificialanalysis.ai/models/quasar-438b/providers); AI Act dates from the [European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai).*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.