Multiverse's Quasar 438B Claims Speed, but Real-World Agent Performance Remains Unclear
The Spanish AI company unveiled its first large-scale model, a compressed 438-billion-parameter system designed for coding and enterprise agents, but questions linger about whether benchmark speeds translate to practical deployment.
Multiverse Computing unveiled Quasar 438B on Wednesday, a compressed reasoning model with 438 billion parameters that the Spanish firm built specifically for coding and enterprise agents. The model achieves a score of 43 on Artificial Analysis' Intelligence Index and 69.3 on Terminal-Bench v2.1, with output speeds measured at roughly 183 tokens per second. According to Multiverse, Quasar ranks as the highest-scoring European model on the Intelligence Index, surpassing Mistral Medium 3.5 (which scored 30) and NVIDIA Nemotron 3 Ultra (which scored 38).
The model occupies a distinctive position in the market. It falls short of leading systems in coding performance, 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. Quasar supports a context window of one million tokens, operates in both English and Spanish, and runs through the Multiverse CompactifAI API.
Compression claims, missing details
The foundation of Multiverse's approach rests on CompactifAI, a technology designed to compress large language models so they consume less memory and computational resources. According to the company, it can shrink model size by 80% to 95% while incurring only minimal accuracy loss. However, Multiverse has not revealed the degree to which Quasar underwent compression or identified the original model from which it derived.
In July, Multiverse secured $570 million in Series C funding to expand its compressed model library and bring the technology to market. Quasar represents the most significant commercial application of this strategy to date.
Critical specifications remain undisclosed. Multiverse has not specified what hardware infrastructure is necessary to operate Quasar or quantified the memory and compute savings that compression delivers. For agent applications, which frequently invoke the model multiple times alongside other tools during task execution, these details carry substantial weight.
Coding benchmarks show tradeoffs
On Terminal-Bench v2.1, Quasar's score of 69.3 positions it above Mistral Medium 3.5 but considerably behind leading frontier systems. Claude Opus 5 leads the benchmark with a score of 89.1.
Multiverse targets Quasar toward software engineering, technical copilots, research, and workflow automation use cases. The one-million-token context window permits agents to manage substantial codebases and retain information as tasks advance, though larger context sizes demand additional computational power. This capability proves particularly valuable in software development scenarios, where code that satisfies all tests may still confuse subsequent AI agents if prior context gets lost.
Agent latency beyond throughput
Artificial Analysis measured Quasar's time to first token at approximately 1.1 seconds and its ability to generate a 500-token response, including reasoning steps, at around 15.3 seconds. While these figures demonstrate respectable performance, agent systems face additional latency sources. Those numbers are fast, but an agent also has to wait for tools, process growing context, and make repeated model calls over the course of a task. The tooling infrastructure supporting agents continues to mature and adapt to model requirements.
Proprietary model, open questions
Quasar remains proprietary and accessible solely through Multiverse's API, preventing developers from examining model weights or deploying it on independent hardware. This restriction currently makes it difficult to verify whether Multiverse's reported speed metrics will persist in practical agent deployments.
Quasar arrives amid a broader trend in which European AI companies construct proprietary model and compute infrastructure rather than depending on American hyperscalers. Multiverse has selected a distinct path, leveraging compression to render a 400-billion-plus-parameter model more economical and efficient to operate. The coming months will reveal whether this compression-focused strategy can withstand real-world performance scrutiny.
Source: The New Stack