Pulsar 16B debuted as an open reasoning model with 16.15 billion parameters but only 3.1 billion active ones, which Multiverse says delivers performance comparable to 30B-class models.
On an NVIDIA Blackwell GPU, the model reached 4,808 tokens per second at 32 concurrent requests—up 43% from 3,363 for the base model—while cutting time-to-first-token to 1.24 seconds from 2.18.
Multiverse said Pulsar 16B beat gpt-oss-20B on nearly every benchmark, including a 15-point advantage on AIME, while preserving instruction following, tool use and 100,000-token retrieval performance.
Built on NVIDIA's Nemotron 3 Nano and compressed with Multiverse's CompactifAI plus NVIDIA optimization tools, the model is aimed at local or single-node enterprise deployments with tighter GPU memory limits.
The Apache 2.0-licensed model is now available on Hugging Face, underscoring a push to make higher-end reasoning models cheaper to run outside cloud-scale infrastructure.
Can a new quantum-inspired AI model solve the industry's crippling memory shortage?
As AI models shrink to think faster, what crucial knowledge might they be leaving behind?
Pulsar 16B: 30B-Class Reasoning at Half the Parameters—A New Benchmark for Efficient AI Models
Overview
Multiverse Computing has launched Pulsar 16B, a new large language model that sets a new standard for efficiency in AI reasoning. Pulsar 16B matches the reasoning power of leading 30B-class models but uses only 16 billion parameters, making it much more efficient. This breakthrough addresses the market’s need for powerful yet resource-friendly AI solutions. Traditionally, deploying large language models required significant computational resources, but Pulsar 16B overcomes these constraints, enabling advanced AI capabilities on less powerful hardware. As a result, it makes sophisticated AI more accessible and practical for a wider range of real-world applications.