Spanish Multiverse Computing has released Quasar 438B — its first large language model in the 438 billion parameter class, focused on enterprise agents and coding. On the independent Artificial Analysis Intelligence Index v4.1.1, it scores 43 points — the best result among European models, and its most demonstrable strength is long context: 75.0 points on AA-LCR with a 1M token window, on par with Grok 4.6 high. The model is proprietary, operates in English and Spanish, and is available via the CompactifAI API at $0.60 per million input tokens and $1.80 per million output tokens. Weak spots include agentic coding, where Quasar 438B lags the leader on Terminal-Bench v2.1 by nearly 20 points, and increased verbosity, which at current prices noticeably affects the final bill.
What Happened
Multiverse Computing from Spain introduced Quasar 438B — its first proprietary model in the large language model class with 438 billion parameters, designed for enterprise agents and coding. The model is proprietary, supports a context of up to 1M tokens, and operates in English and Spanish. According to Artificial Analysis, on the Artificial Analysis Intelligence Index v4.1.1, Quasar 438B scores 43 points, outperforming Mistral Medium 3.5 with 30 points and NVIDIA Nemotron 3 Ultra with 38. On the long-context benchmark AA-LCR, the result is 75.0 points, on par with Grok 4.6 high. In agentic coding on Terminal-Bench v2.1, the model scores 69.3 points compared to 89.1 for the leader, Claude Opus 5. Speed estimates in the sources differ: independent measurements by Artificial Analysis show 176 tokens/s with a TTFT of 1.06 seconds, while the release materials mention a response "in 15.3 seconds for 500 tokens." The model is available via the CompactifAI API at a price of $0.60 per million input tokens and $1.80 per million output tokens; registration is open at dashboard.compactif.ai.
Context
Multiverse Computing has historically been associated with the model compression technology CompactifAI — the same name is used for the API through which the company released its first large model. Prior to this release, Spain did not have a notable player in the class of large reasoning models, and the European benchmark in this category was considered to be Mistral with its mid-tier model. This leads to the main caveat regarding the phrase "Europe's Leading AI Model" in the release title: it relies on a comparison with a regional set of models, not the frontier. The asymmetry of Quasar 438B's profile is indicative: in long context, it is indeed comparable to frontier systems, while in agentic coding it remains a follower. This is a typical profile of a specialized model with a strong niche, not a universal flagship, and this is how this release should be read. A separate open question is the architecture: the connection between Quasar 438B and CompactifAI's compression technology is not disclosed in available sources.
Why This Matters for the Industry
For the industry, this is Spain's first notable entry into the class of large reasoning models: Quasar 438B outperforms the European flagship Mistral Medium 3.5 by 13 points on the intelligence index with fewer parameters, which strengthens the thesis of "sovereign AI" in Europe and the emergence of independent options outside the US and China. The release's economics put pressure on the agent API market: input tokens at $0.60 per million with a speed of 176 tokens/s and a TTFT of 1.06 seconds make document-heavy agent pipelines significantly cheaper and reduce the value of closed access to long context as a competitive advantage. Cheap input allows for large-scale runs — monitoring document streams and bulk processing — without a heavy RAG layer; if the price holds and verbosity is reduced, the pattern of "an agent without RAG over a medium corpus" could become the default for MVPs, and long-context APIs could become an independent product category with price competition. Increased model routing is also likely: a cheap model for reading and long context will be combined with a frontier model for coding, where Quasar 438B currently lags significantly. The key test in the coming months — whether a technical report, SLA, enterprise terms, and real deployment cases will appear: without them, Multiverse will remain in the category of regional providers, with them — it may occupy the niche of a budget option for long context in the European API market.
Why This Matters for Users
If you are building agents or copilots, the model is available without your own infrastructure: via the CompactifAI API with registration at dashboard.compactif.ai, so a pilot can be launched immediately and on your own evals you can check actual verbosity, the cost of a single task, and latency under your load. A practical first scenario is multi-document analysis and questions on long documents and research tasks in English and Spanish: here the model is strong (75.0 on AA-LCR with a 1M token window), and input at $0.60 per million tokens allows for cheap large-scale runs. For products where the core is coding or universal agents, it is too early to migrate: in terminal coding, the gap with the frontier is nearly 20 points. An important practical nuance is verbosity: an approximately fivefold excess over the median multiplies the bill for output tokens at $1.80 per million and the agent's runtime, so the final cost of a task may be significantly higher than the cheap input promises. The model officially operates in English and Spanish; nothing is said in the sources about Russian language support, so Russian-speaking teams should pilot through English and check quality on their own data.
What Is Still Unknown / Limitations
The title "Europe's Leading AI Model" is an interpretation, not an established fact: it relies on a comparison with the mid-tier model Mistral Medium 3.5, not the frontier. AA-LCR 75.0 is one benchmark from one evaluator (Artificial Analysis); the methodology and run protocols are not disclosed in available sources, and the claimed long-context strength has not yet been confirmed by independent reproductions. There is an internal contradiction regarding speed: 15.3 seconds for 500 tokens (about 33 tokens/s) versus 176 tokens/s according to Artificial Analysis — likely, these are different metrics (end-to-end task latency versus pure generation speed), but the methodology is not disclosed anywhere. Verbosity approximately 5 times higher than the median can eat up the benefit of cheap input at a price of $1.80 per million output tokens. There is no data on SLA, rate limits, uptime, and production cases in the sources, so reliability for production is not proven. There is no technical report, and the connection between Quasar 438B's architecture and CompactifAI compression is not disclosed. Nothing is stated about Russian language support.