Overview
Quasar 438B is the first large model released by Multiverse Computing, launched on 2 September 2026. The company describes it as "our flagship reasoning model, built for enterprise-scale agents and coding"; it runs in English and Spanish and carries 438 billion parameters.
Multiverse positions the model on the Artificial Analysis Intelligence Index v4.1.1, a composite of nine evaluations — GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR. Quasar scores 43 there, which the company reports as the highest result of any European model in that comparison, ahead of Mistral Medium 3.5 at 30, NVIDIA Nemotron 3 Ultra at 38 and Inkling at 42, in a field led by Claude Opus 5 at 63.
Speed is the other half of the pitch: Multiverse says Quasar returns 500 tokens, thinking time included, in 15.3 seconds, against 18.8 seconds for Mistral Medium 3.5, 25.7 for Nemotron 3 Ultra and 48.3 for Inkling. Only three models in its comparison answer faster — Nemotron 3.5 Lightning at 9.4s (index 24), Gemini 3.5 Flash-Lite at 10.8s (37) and Gemini 3.7 Flash at 11.5s (56) — and only the last of those also scores higher.
The model is served through the CompactifAI API, Multiverse's hosted platform for its compressed and specialised models. It is exposed as model id `quasar-438b` on an OpenAI-compatible `POST /v1/chat/completions` endpoint with tool/function calling and structured output (`response_format`). Reasoning is always on and cannot be disabled; `reasoning_effort` accepts `high` or `max`, with `max` the default. Published pricing is $0.60 per million input tokens and $1.80 per million output tokens.
Multiverse Computing is headquartered in Donostia-San Sebastián, Spain, with offices in the United States, Canada and across Europe. Its stated focus is "sovereign and efficient AI" — models organisations can deploy inside their own infrastructure — and its CompactifAI line is built on model compression.
| Released | 2026-09-02 |
|---|---|
| License | Proprietary (hosted CompactifAI API) |
| Weights | API only |
| Parameters | 438 billion |
| Modalities | Text |
| Status | Generally available |
Benchmarks
Quasar 438B against the field, transcribed from Multiverse Computing's launch post (2 September 2026). Blank cells were not reported there.
| Benchmark | Quasar 438B | Claude Opus 5 | Inkling | NVIDIA Nemotron 3 Ultra | Mistral Medium 3.5 |
|---|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.1.1 | 43 | 63 | 42 | 38 | 30 |
| AA-LCR (long-context reasoning) | 75 | 75.7 | — | — | — |
| Terminal-Bench v2.1 | 69.3 | 89.1 | — | — | — |
| Time to return 500 tokens (thinking included) | 15.3 s | — | 48.3 s | 25.7 s | 18.8 s |
This model's scores
- Artificial Analysis Intelligence Index v4.1.143
- AA-LCR (long-context reasoning)75
- Terminal-Bench v2.169.3
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.60 / 1M tokens |
|---|---|
| Output | $1.80 / 1M tokens |
CompactifAI API list price for model id quasar-438b. Billing is usage-based with no monthly commitment or minimum fee.
Strengths
- 43 on the Artificial Analysis Intelligence Index v4.1.1 — the highest European result in Multiverse's published comparison
- Fast for a reasoning model: 500 tokens including thinking time in 15.3 seconds, versus 18.8s for Mistral Medium 3.5 and 48.3s for Inkling
- 75.0 on AA-LCR for long-document reasoning — level with Grok 4.6 (high) and within a point of Claude Opus 5 at 75.7
- 69.3 on Terminal-Bench v2.1, leading Mistral Medium 3.5 by 18.7 points and Nemotron 3 Ultra by 15.4
- Cheap for its tier at $0.60 / $1.80 per million tokens on the CompactifAI API
- OpenAI-compatible endpoint with tool calling and structured output, so existing clients work unchanged
Best for
- Reach for it for enterprise agent workloads that need reasoning without frontier-model latency
- Reach for it for coding and terminal-driven automation where Terminal-Bench-style task completion matters
- Reach for it for long-document analysis — extracting and connecting facts spread across large inputs
- Reach for it for Spanish-language work, one of the two languages the model is built for
How to access
| Provider | Model ID |
|---|---|
| CompactifAI API ↗ | quasar-438b |
FAQ
What is Quasar 438B?
Quasar 438B is a 438-billion-parameter reasoning model launched by Multiverse Computing on 2 September 2026 — the company's first large model. It runs in English and Spanish and is aimed at enterprise-scale agents, coding and multi-step work. It is served through the CompactifAI API.
How does Quasar 438B score on benchmarks?
Multiverse reports 43 on the Artificial Analysis Intelligence Index v4.1.1, 75.0 on AA-LCR and 69.3 on Terminal-Bench v2.1. In its published comparison that index score leads Mistral Medium 3.5 (30), NVIDIA Nemotron 3 Ultra (38) and Inkling (42), while Claude Opus 5 leads the field at 63.
How fast is it?
Multiverse says Quasar returns 500 tokens, thinking time included, in 15.3 seconds. In its comparison Mistral Medium 3.5 takes 18.8 seconds, Nemotron 3 Ultra 25.7 and Inkling 48.3. Three models answer faster — Nemotron 3.5 Lightning (9.4s), Gemini 3.5 Flash-Lite (10.8s) and Gemini 3.7 Flash (11.5s) — but only Gemini 3.7 Flash also scores higher on the index.
What does Quasar 438B cost, and how do I call it?
CompactifAI lists $0.60 per million input tokens and $1.80 per million output tokens. The model id is quasar-438b on an OpenAI-compatible POST /v1/chat/completions endpoint, with tool/function calling and structured output supported. Reasoning is always enabled; reasoning_effort accepts high or max, and max is the default.
Are the weights open?
No. Quasar 438B is proprietary and available through the hosted CompactifAI API; Multiverse Computing has not published downloadable weights for it.