Multiverse Computing Launches 16.15B-Parameter Pulsar 16B as Throughput Jumps 43%
Updated
Updated · GlobeNewswire · Jun 23
Multiverse Computing Launches 16.15B-Parameter Pulsar 16B as Throughput Jumps 43%
2 articles · Updated · GlobeNewswire · Jun 23
Summary
Pulsar 16B debuted as an open Apache 2.0 reasoning model on Hugging Face, compressing NVIDIA’s 31.6B-parameter Nemotron-3-Nano base to 16.15B parameters without retraining from scratch.
Benchmark results showed the smaller model largely matching its 30B-class base and beating gpt-oss-20B on most measures, including 87.22 on AIME 2025 and 71.41 on GPQA-Diamond.
On an NVIDIA Blackwell GPU with 32 concurrent requests, Pulsar 16B in FP8 reached 4,808 tokens per second versus 3,363 for the base model, while time-to-first-token fell to 1.24 seconds from 2.18.
Multiverse said the lower memory footprint and preserved long-context performance make the model more practical for single-node, on-premises and latency-sensitive enterprise deployments.
Can a compressed 16B model truly rival a 30B model, or are we sacrificing subtle reasoning for speed and efficiency?
As advanced AI moves from big tech clouds to private servers, what new security vulnerabilities and ethical dilemmas will emerge?
How will the rise of national 'Sovereign AI' models reshape global technology competition and digital information borders?
Pulsar 16B and CompactifAI: Quantum-Inspired Model Compression Achieves Superior AI Efficiency and Sustainability in 2026
Overview
On June 23, 2026, Multiverse Computing launched Pulsar 16B, marking a major milestone in artificial intelligence. Pulsar 16B’s standout efficiency comes from CompactifAI, a quantum-inspired engine designed to revolutionize AI model deployment. CompactifAI enables extreme model compression while keeping strong performance, which helps solve key challenges in AI scalability and resource use. By drastically reducing the computational power needed for AI operations, CompactifAI makes advanced AI models more accessible, efficient, and sustainable for a wide range of applications.