The Spanish firm Multiverse Computing introduced Quasar 438B on Wednesday, positioning its compressed 438-billion-parameter reasoning model as a solution for enterprise and coding agents. While the model achieves a 43 score on Artificial Analysis' Intelligence Index and 69.3 on Terminal-Bench v2.1, with reported output speeds around 183 tokens per second, the relationship between benchmark performance and practical agent speed remains murky.
Deploying a model with 438 billion parameters typically conflicts with speed requirements, but Multiverse argues that compression technology can bridge that gap. The company ranks Quasar as Europe's highest-scoring model on the Intelligence Index, surpassing Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38).
Quasar occupies a strategic middle position in the market. It doesn't rival top-tier systems on coding tasks, yet Multiverse is betting that a 438B-parameter model can be fast and inexpensive enough for agents that repeatedly reason, call tools, and check the results. The model supports a one-million-token context window, operates in English and Spanish, and runs through the Multiverse CompactifAI API.
Compression claims, missing details
Multiverse developed CompactifAI specifically to compress large language models, reducing memory and compute overhead. The company claims it can shrink model size by 80% to 95% while maintaining reasonable accuracy levels, yet it has not revealed how much compression was applied to Quasar or which base model served as its starting point.
The firm secured $570 million in Series C funding during July to expand its compressed model portfolio and scale commercialization efforts. Quasar represents the most significant validation of this compression strategy to date.
Critical infrastructure details remain undisclosed. Multiverse has not specified what hardware Quasar requires or quantified how much compression reduces memory and compute demands. For agent workflows that invoke the model repeatedly alongside other tools, these specifics carry substantial weight.
Coding benchmarks show tradeoffs
On Terminal-Bench v2.1, Quasar's 69.3 score exceeds Mistral Medium 3.5 but trails significantly behind leading systems—Claude Opus 5 reaches 89.1 on the same benchmark. Multiverse targets Quasar toward software engineering, technical copilots, research applications, and workflow automation.
The one-million-token context window provides agents substantial capacity for large codebases and information retention across task execution phases. However, processing extended context demands additional computational resources. In coding scenarios especially, this matters: code that passes all tests can still confuse subsequent AI agents if prior context gets lost.
Agent latency beyond throughput
Artificial Analysis measured Quasar's time-to-first-token at roughly 1.1 seconds and its ability to generate a 500-token response (including reasoning) at approximately 15.3 seconds. While these figures appear competitive, an agent also has to wait for tools, process growing context, and make repeated model calls over the course of a task. The broader agent infrastructure ecosystem continues maturing to meet these models' requirements.
Proprietary model, open questions
Quasar remains proprietary and accessible solely through Multiverse's API, preventing developers from examining model weights or executing it on proprietary infrastructure. This limitation currently makes it difficult to verify whether Multiverse's reported speed metrics will translate to production agent deployments.
Quasar emerges amid a broader European push toward building independent AI models and compute infrastructure rather than depending on U.S. cloud providers. Multiverse has chosen a distinct path: leveraging compression to make a 400-billion-plus-parameter model economically and operationally viable. Whether this approach withstands real-world scrutiny will determine its market viability.
Source: The New Stack