Quasar 438B is on Eden AI: Benchmarks, Pricing, and API Access

On September 2, 2026, Multiverse Computing released Quasar 438B. It scores 43 on the Artificial Analysis Intelligence Index, the highest result of any European model, and ranks 13th out of 178 models measured worldwide. It ships with a 1 million token context window, 69.3 on Terminal-Bench v2.1 for agentic coding, and 75.0 on AA-LCR for long-context reasoning, at 0.60 dollars per million input tokens and 1.80 per million output.

Two days later, it is live on Eden AI. You can call it today with the model string compactifai/quasar-438b.

What makes this release interesting is not only the score. It comes from a company almost nobody had on their European shortlist, and it is an efficiency result as much as a capability one: Quasar beats NVIDIA's Nemotron 3 Ultra while carrying 112 billion fewer parameters. That is exactly what you would expect from a team that spent six years doing nothing but making models smaller.

Multiverse Computing report that Quasar 438B scores 43 on the Artificial Analysis Intelligence Index, the highest result of any European model evaluated, a figure French tech press has since repeated.

‍

Key Takeaways

• Quasar 438B scores 43 on the Artificial Analysis Intelligence Index, the highest of any European model and 13th of 178 worldwide.

• It is built for reasoning, agentic coding and long-context work: 69.3 on Terminal-Bench v2.1, 75.0 on AA-LCR, and a 1 million token context window.

• Pricing is 0.60 dollars per million input tokens and 1.80 per million output, well below the US frontier models it gets benchmarked against.

• It is live on Eden AI now as compactifai/quasar-438b, callable with the same OpenAI-compatible request shape you already use.

• Reasoning is always on, and Eden AI returns the reasoning trace separately in reasoning_content, so you can inspect how the model reached its answer.

‍

What's New in Quasar 438B

A Compression Company That Built Its Own Frontier Model

Multiverse Computing was founded in San Sebastián in 2019 by Enrique Lizaso, Román Orús, Alfonso Rubio-Manzanares and Sam Mugel. For six years their product was CompactifAI, a model compression platform built on tensor network methods borrowed from quantum physics, used to cut inference costs by 50 to 80 percent for customers including BBVA, Bosch, BASF, Crédit Agricole CIB, Repsol and the Bank of Canada.

That is an unusual place to start from, and it shows in the result. Quasar 438B is the company's first large model of their own, released after a 215 million dollar Series B in 2025 backed by HP Tech Ventures, Toshiba, Forgepoint Capital International, Bullhound Capital and Santander Climate VC. The company now employs around 480 people.

The efficiency thesis is visible in the numbers. Quasar reaches an Intelligence Index of 43 with 438 billion parameters. NVIDIA's Nemotron 3 Ultra reaches 38 with 550 billion. Five points more capability from 112 billion fewer parameters is a compression story, and it suggests the CompactifAI work fed directly into how Quasar was built rather than sitting alongside it.

A 1 Million Token Context Window

Quasar ships with a 1 million token context window, which puts it in the same bracket as the largest US frontier models and ahead of every other European-built model. For agents this is the specification that changes what you can attempt: an entire mid-sized repository, a full contract set, or a long research corpus fits in a single call.

Context size on its own is not enough, which is why the AA-LCR score below matters more than the window figure.

Always-On Reasoning with Effort Controls

Quasar is a reasoning model, and reasoning is always active. You control how much of it happens through the reasoning_effort parameter, which accepts high and max, with max as the default.

Through Eden AI, the reasoning trace comes back in a separate reasoning_content field on the response, distinct from the answer in content. That is genuinely useful in production: you can log the reasoning for audit or debugging without it polluting what you show the user.

Tool Calling and Structured Outputs

Both are supported. Tool calling and response_format structured outputs are the baseline requirements for anything agentic, and Quasar covers them, which is why it holds up on Terminal-Bench rather than only on chat benchmarks.

Benchmark Results

Intelligence Index: 43, the Highest of Any European Model

Artificial Analysis is an independent evaluation company, so this is a third-party result rather than a vendor claim. Their Intelligence Index is a composite across reasoning, knowledge and agentic evaluations, including GPQA Diamond, Humanity's Last Exam, Terminal-Bench, SciCode, AA-LCR and others.

Quasar scores 43. For context:

• Quasar 438B: 43

• NVIDIA Nemotron 3 Ultra: 38, with 550 billion parameters

• Mistral Medium 3.5: 30

Quasar ranks 13th out of 178 models measured, which is worth stating precisely: 13th in the world, not 13th in Europe.

Terminal-Bench v2.1: 69.3 for Agentic Coding

Terminal-Bench measures whether a model can complete real tasks in a terminal environment, which is a much better proxy for agent work than single-turn coding benchmarks. 69.3 puts Quasar in credible territory for coding agents rather than only for code suggestions.

AA-LCR: 75.0 for Long-Context Reasoning

AA-LCR tests reasoning quality across long documents, not just whether the model accepts a long input. This is the score to read alongside the 1 million token window, and 75.0 paired with 1M context is the standout combination in the European field.

Speed

Multiverse report a 500 token response, reasoning included, in 15.3 seconds, at 176 tokens per second. Their framing is that only three models in the Artificial Analysis comparison respond faster, and only one of those scores higher on intelligence.

[H2] Pricing and Cost Per Task

[H3] Base Pricing

• Input: 0.60 dollars per million tokens

• Output: 1.80 dollars per million tokens

That is a fraction of what the US frontier models it gets compared to charge, which is the core of the commercial argument. Multiverse are also running a launch promotion with the first month free on their own platform, so treat that as promotional rather than as the standing rate.

A Real Worked Example

Because reasoning is always on, the output token count on a Quasar call includes the reasoning tokens, not just the words you see. This is the single most important thing to understand before you budget for it, so here is an actual measured call through Eden AI rather than an estimate.

Prompt: "In one sentence, what is an AI gateway?"

• Prompt tokens: 22

• Completion tokens: 919

• Cost: 0.0017 dollars

The visible answer was one sentence. The other 880 or so completion tokens were reasoning, returned separately in reasoning_content. The model genuinely thought about the question, and you paid for that thinking.

The practical consequence is straightforward and it is the same rule that applies to every reasoning model: do not put a reasoning model on trivial steps. At roughly 0.0017 dollars for a one-sentence answer, a thousand such calls costs about 1.70 dollars, which is nothing if those thousand calls are hard reasoning tasks and wasteful if they are date extraction.

Where It Sits on Price

Against the models it gets benchmarked against, Quasar is inexpensive. Against open-weight models in a similar intelligence band, several of which you can run on European infrastructure, it is priced as a premium product. Artificial Analysis make this point in their own analysis.

That is not a criticism of the pricing, it is a routing instruction. Quasar earns its cost on the steps that need a frontier-adjacent European model, and something cheaper should handle the rest. The section below shows how.

Quasar 438B Best Use Cases

Agentic Coding and Software Engineering

This is the strongest fit. 69.3 on Terminal-Bench with tool calling and structured outputs is the combination coding agents actually need, and the 1 million token window means the agent can hold a real codebase rather than a handful of files.

Long-Context Document and Repository Work

The 1M window plus 75.0 on AA-LCR is the pairing to look for here. Contract review across a full document set, research synthesis over a corpus, or whole-repository analysis are all workloads where the long-context score matters more than the headline intelligence figure.

Spanish-Language Products

Quasar supports English and Spanish. For teams building for Spanish-speaking markets, that focus is an advantage rather than a limitation: this is the strongest European-built model available and it was trained with Spanish as a first-class language, not as an afterthought in a list of thirty.

If your product serves French, German or Italian users, pick a model with broader European language coverage. Mistral remains the better default there, and both are available through the same Eden AI key.

Multi-Step Enterprise Automation

Multiverse position Quasar for enterprise-scale agents and complex multi-step workloads, and their existing customer base in banking, industry and the public sector is consistent with that. Long workflows with tool use and audit requirements are a natural fit, particularly given the separate reasoning trace.

When to Use Quasar 438B

The pattern that works with any reasoning model is tiered routing: route each step to the cheapest model that can do it, and reserve the expensive one for the steps that need it.

A sensible three-tier split for a European stack:

• Cheap tier, for classification, extraction, formatting and routing decisions: an open-weight model hosted in Europe, on Scaleway, OVHcloud or IONOS.

• Standard tier, for general generation and multilingual work: Mistral Medium 3.5 or Mistral Large 3.

• Hard tier, for the reasoning, planning and agentic coding steps: Quasar 438B.

In most agent workloads the hard tier is a small share of calls and a large share of value, which is exactly the shape that makes tiered routing pay.

How to Call Quasar 438B on Eden AI

Quasar 438B is available on Eden AI as compactifai/quasar-438b, on the OpenAI-compatible chat completions endpoint. If you already call any other model through Eden AI, this is a one-line change.

import requests

headers = {"Authorization": "Bearer YOUR_API_KEY"}
payload = {
    "model": "compactifai/quasar-438b",
    "messages": [
        {"role": "user", "content": "Refactor this function and explain your changes."}
    ],
}

r = requests.post(
    "https://api.edenai.run/v3/llm/chat/completions",
    json=payload,
    headers=headers,
)
data = r.json()

answer = data["choices"][0]["message"]["content"]
reasoning = data["choices"][0]["message"]["reasoning_content"]

‍

To route the same request to another model, change the model value and nothing else. Multiverse Computing's other models are available too, including compactifai/carina-60b and compactifai/hypernova-60b, alongside 800 models from 70+ providers on the same key.

Why This Matters for European AI

Until this release, the European conversation had a gap in it. You could get EU jurisdiction, a signable DPA and data residency, but at the top of the capability range you were still reaching for a US model. Quasar does not close that gap entirely, Claude Opus 5 scores 63 on the same index, but it narrows it meaningfully and at a price the frontier cannot match.

It also broadens the map. European AI has been a France and Germany story for two years. The highest-scoring European-built model is now Spanish, from a company that got there by a completely different route, and that is a healthier picture for the continent than a single national champion.

‍