Pulsar 16B: Frontier AI Reasoning at Half the Infrastructure Footprint

DONOSTIA, Spain – June 23, 2026 – The race for artificial intelligence dominance has long been a game of scale, where bigger models meant better performance, often at staggering computational and financial costs. Today, Multiverse Computing, in a deep collaboration with NVIDIA, challenges that paradigm with the launch of Pulsar 16B, an open-source reasoning model that delivers the power of a 30-billion-parameter model at roughly half the size.

This release signals a potential shift in the AI landscape, moving the focus from sheer scale to intelligent efficiency. By making high-end reasoning capabilities accessible without the need for massive cloud infrastructure, Pulsar 16B opens the door for enterprises, especially in regulated industries, to deploy powerful AI on their own terms.

The Efficiency Revolution: Redefining AI's Footprint

The central claim of Pulsar 16B is its ability to punch far above its weight class. With just 16.15 billion parameters, the model demonstrates reasoning and knowledge capabilities on par with its 31.6B-parameter base model, NVIDIA's Nemotron-3-Nano-30B. This isn't a case of minor optimization; it's a significant reduction in model size without a corresponding drop in intelligence.

Validation across a suite of demanding industry benchmarks underscores this achievement. On the AIME 2025 math reasoning test, Pulsar 16B scored 87.22, nearly identical to its uncompressed parent model and a full 15 points ahead of the comparable gpt-oss-20B model. The gap is even more pronounced in the GPQA-Diamond benchmark, a set of PhD-level science questions designed to be 'Google-proof.' Here, Pulsar 16B matches the 30B model's score of 71.41, leaving gpt-oss-20B far behind at 58.88.

The magic behind this compression is Multiverse Computing’s proprietary 'CompactifAI' technology. Rather than relying solely on traditional methods like pruning (removing connections) or quantization (reducing numerical precision), which can degrade performance when applied aggressively, CompactifAI employs quantum-inspired mathematics. Specifically, it uses tensor networks to identify and eliminate mathematical redundancy within the trained neural network, preserving the complex reasoning pathways learned during the model's initial training. This allows it to maintain critical functions like instruction-following and tool-use, where other compressed models often falter.

The practical benefits are immediate. On an NVIDIA Blackwell GPU, Pulsar 16B in its FP8 precision format delivers 43% higher throughput than its base model and slashes the time-to-first-token—a key metric for user experience in interactive applications—by over 40%. For businesses running document-intensive workloads, the model’s long-context performance remains robust, with near-perfect information retrieval even beyond 100,000 tokens.

A Strategic Alliance for Accessible AI

Pulsar 16B is the product of a deep, strategic alignment between Multiverse Computing, a member of the NVIDIA Inception program, and the chip-making giant itself. This is not a simple case of one company using another's hardware; it's a co-engineered effort to advance the entire AI ecosystem. The model is built directly upon NVIDIA's Nemotron architecture—a state-of-the-art Hybrid Mamba2-Transformer with a Mixture-of-Experts (MoE) design—and was compressed using a workflow that integrates NVIDIA's own Model Optimizer and Megatron Bridge libraries.

This collaboration showcases NVIDIA's broader strategy: fostering a vibrant ecosystem of highly optimized models that maximize the performance of its accelerated computing platforms. By providing the architectural foundation and the optimization tools, NVIDIA empowers partners like Multiverse to create specialized, efficient models that cater to specific market needs. The performance of Pulsar 16B was, in turn, independently validated on NVIDIA's own accelerated infrastructure, lending significant credibility to the claims.

By making the model available on Hugging Face under the permissive Apache 2.0 license, the collaboration also makes a significant contribution to the open-source community. It provides researchers and developers with a powerful, frontier-grade reasoning engine that is far more accessible to run and fine-tune than its larger predecessors, sparking a new wave of innovation on a more economical foundation.

Empowering the Sovereign Enterprise

Perhaps the most significant impact of Pulsar 16B will be in the realm of 'sovereign AI.' For enterprises in highly regulated sectors like finance, healthcare, energy, and government, the public cloud is not always a viable option for sensitive workloads. Data residency laws, privacy regulations like GDPR, and the existential risk of data breaches demand a level of control that only on-premise or private cloud deployments can provide.

Until now, running powerful AI locally meant a compromise. “Running advanced AI locally has historically required compromising on model size or performance,” said Enrique Lizaso, cofounder and CEO of Multiverse Computing. “What we’re demonstrating with Pulsar 16B is that frontier-grade reasoning can now be deployed without the overhead of cloud-scale infrastructure, at a footprint enterprises can actually run and scale economically.”

Pulsar 16B's reduced memory and computational footprint directly addresses this challenge. It makes single-node deployments, latency-sensitive systems, and high-concurrency agentic workflows feasible within an organization's own secure perimeter. This allows companies like Multiverse's existing clients—which include Iberdrola, Bosch, and the Bank of Canada—to leverage advanced AI while maintaining full sovereignty over their data, governance, and compliance posture.

Navigating the Competitive Open-Source Landscape

Pulsar 16B does not enter the market in a vacuum. The AI industry is rapidly pivoting towards efficiency, with a host of powerful, compact models like Mistral 7B, Google's Gemma family, and Claude 3 Haiku gaining traction. Multiverse Computing has positioned Pulsar 16B to compete directly with models in the 20B-parameter class, most notably taking aim at OpenAI's gpt-oss-20B and outperforming it across a range of production-critical benchmarks.

This release builds on Multiverse's established track record, including its previous launch of HyperNova 60B, which offered a more compact alternative to 120B-parameter models. Bolstered by a recent €189 million Series B funding round specifically for its CompactifAI technology, the Spanish company is cementing its role as a key innovator in the field of AI efficiency.

By combining the architectural prowess of NVIDIA with its own unique compression science, Multiverse Computing has delivered a model that is not just smaller, but smarter about its size. For a world grappling with the immense energy and hardware demands of AI, Pulsar 16B offers a compelling vision of a more sustainable and accessible future for artificial intelligence.