AI
Multiverse's 438B Model Promises Speed for AI Agents, but Benchmarks Reveal Mixed Results
The Spanish AI company claims its Quasar model delivers the performance needed for agent workflows, though independent testing suggests the real-world gains remain uncertain.
The New Stack
Multiverse Computing, a Spanish AI firm, introduced Quasar 438B this week—a massive reasoning model built to serve coding and enterprise agents. The company argues that model compression technology can make a 438-billion-parameter system practical for repeated reasoning cycles, tool calls, and result verification.
The Compression Gamble#
Multiverse's CompactifAI compression system claims to shrink models by 80–95% while retaining most of their capabilities. Quasar is the largest real-world test of this approach, following the company's $570 million Series C funding in July. However, Multiverse has yet to reveal how much Quasar itself was compressed or which base model it derived from. The company also hasn't specified what hardware infrastructure runs the model or the actual memory and compute savings agents would experience when deploying it repeatedly.
Performance in Context#
On Artificial Analysis' Intelligence Index, Quasar scores 43—outpacing Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38), making it Multiverse's claim for Europe's top-ranked model by that metric. For coding tasks, it achieves 69.3 on Terminal-Bench v2.1, trailing Claude Opus 5's 89.1 but ahead of Mistral Medium 3.5. The model produces output at roughly 183 tokens per second and takes about 1.1 seconds to start responding.
While these figures appear competitive, the full picture remains murkier. A 500-token response including reasoning takes approximately 15.3 seconds—respectable, but agents must also account for tool latency, growing context windows, and multiple model calls across a single workflow.
Missing Transparency and Real-World Unknowns#
Quasar remains proprietary and accessible only through Multiverse's API. Developers cannot inspect model weights or run it on their own infrastructure, making it difficult to validate whether reported speeds translate to actual agent deployments at scale. The one-million-token context window theoretically helps agents maintain coherence across large codebases and complex tasks, though wider contexts demand additional compute.
Quasar represents a distinct strategy as European AI companies increasingly build homegrown models and infrastructure rather than relying on U.S. cloud providers. Multiverse is betting compression—rather than scale or open-source accessibility—unlocks the cost and speed advantage the market demands. Whether that bet holds up in production remains the central unanswered question.
