Multiverse Computing has launched Pulsar 16B, a new open-source reasoning model built on Nvidia's Nemotron architecture. The model uses 16.15 billion parameters but activates only about 3.1 billion at any time, delivering efficiency the company says rivals larger models.
This efficiency is driven by Multiverse's proprietary CompactifAI compression technology.
Pulsar 16B reports significant performance gains on Nvidia's Blackwell GPUs, with system throughput reaching 4,808 tokens per second-a 43% increase. It reduces the time to first token to 1.24 seconds, a major improvement for high-volume enterprise queries.
The model is available in three precision formats and supports hybrid reasoning methods. It is released under the permissive Apache 2.0 license on Hugging Face, allowing broad commercial use.
Multiverse Computing, based in Spain, raised $215 million in its Series B round in June 2025 to scale its platform. The company has applied quantum computing principles in projects with institutions like the Bank of Canada.
For investors, the model carries no crypto token or associated financial mechanisms. Its open licensing offers an advantage for integration into crypto-native AI projects, contrasting with the usage restrictions of some competing models.
The model's optimization for Nvidia's ecosystem highlights the growing competition in efficient open-source AI, where rivals like Mistral and DeepSeek also operate.