Multiverse Computing just dropped Pulsar 16B, an open-source reasoning model that punches well above its weight class. Built on Nvidia’s Nemotron architecture, the model packs 16.15 billion parameters but only activates about 3.1 billion of them at any given time, delivering performance that the company says is competitive with models twice its size.

The trick behind that efficiency has a name: CompactifAI, Multiverse’s proprietary compression technology.

The numbers behind the hype

Pulsar 16B achieves a system throughput of 4,808 tokens per second, which Multiverse says represents a 43% increase when running on Nvidia’s Blackwell GPUs. That throughput figure accounts for handling multiple concurrent requests, not just a single user asking it to write a haiku about blockchain.

Time-to-first-token dropped to 1.24 seconds from 2.18 seconds. At enterprise scale, where hundreds or thousands of queries hit the system simultaneously, shaving nearly a full second off initial response time is meaningful.

The model ships in three precision formats: BF16, FP8, and NVFP4, representing different trade-offs between computational accuracy and speed. It supports both step-by-step and direct-response hybrid reasoning methods.

Pulsar 16B is released under the Apache 2.0 license and available on Hugging Face, allowing commercial use, modification, and distribution.

Who is Multiverse Computing

Founded in 2019 and headquartered in San Sebastián, Spain, Multiverse Computing raised $215 million in a Series B round in June 2025, capital earmarked to scale the CompactifAI platform. The firm’s track record includes partnerships with institutions like the Bank of Canada, where it has applied quantum computing principles to financial simulations.

What this means for crypto and AI investors

There is no crypto token attached to Pulsar 16B. No airdrop, no governance vote, no liquidity pool.

The open-source licensing is relevant for crypto-native AI projects. Apache 2.0 means anyone can integrate Pulsar 16B without navigating restrictive terms. Compare that to Meta’s Llama models, which carry usage restrictions above certain user thresholds, or proprietary offerings from OpenAI and Anthropic that come with API costs and no ability to self-host modifications.

Multiverse’s decision to optimize specifically for Blackwell GPUs and to validate on Nvidia’s accelerated infrastructure demonstrates the depth of that ecosystem. The competitive landscape for efficient open-source reasoning models includes Mistral, Qwen, and DeepSeek, all of which have released models targeting similar efficiency benchmarks. What separates Pulsar 16B is the compression approach, using quantum-inspired mathematics to achieve parameter efficiency rather than conventional distillation or pruning techniques.