Pick Multiverse Computing if inference cost or hardware limits are blocking a model you already like, since it shrinks open models to fit smaller GPUs, CPUs and devices. Skip it if you only call hosted frontier APIs and never self-deploy. New API users get a free first month, and Comet 60B launched on October 15, 2026.
Multiverse Computing is a San Sebastián company, founded in 2019, whose CompactifAI platform applies tensor networks from quantum physics to compress large language models by a claimed 80-95%. It publishes compressed models built on NVIDIA Nemotron, including Pulsar 16B and Comet 60B, and targets edge devices, private data centers and sovereign AI.
Founded: 2019 · HQ: San Sebastián, Spain · CEO: Enrique Lizaso · Funding: Series C announced July 27, 2026 ($570M, EUR 500M); total funding expected to reach about $800M · Valuation: $1.7B pre-money (Series C, July 2026)
About Multiverse Computing
Multiverse Computing is a Spanish AI infrastructure company founded in 2019 in San Sebastián, Basque Country. Co-founder Enrique Lizaso is chief executive and co-founder Dr. Román Orús is chief scientific officer. Its core product, CompactifAI, applies tensor networks, a mathematical framework from quantum physics, to restructure the weight matrices of large language models into compressed forms. The company says this cuts model size by 80-95% with immaterial accuracy loss, which differs from quantization because the architecture changes rather than only the numerical precision. Its homepage claims 50-80% lower inference costs and up to 2x faster inference.
The company works closely with NVIDIA. Pulsar 16B, released in June 2026, is a compressed derivative of Nemotron 3 Nano 30B, and Hokai lists it as Pulsar 16B. On October 15, 2026 it launched Comet 60B, a compressed version of NVIDIA's Nemotron 3.5 Super VL. Comet halves the total parameters from 120B to 60B, keeps the same 12B active parameters and all three modalities (text, image and video), and ships in BF16, FP8 and NVFP4 formats. The company says accuracy stays within 5% of the original and improves on some benchmarks.
CompactifAI is sold as an API, and the homepage offers new API users a free first month. The platform also includes the CompactifAI Router, which the company says decides in real time whether a workload runs locally or goes to the cloud. Deployment options listed on the site are private cloud, on-premises and edge devices. Recent announcements include a Qualcomm collaboration on efficient models for data centers, a claim that all CompactifAI models now run on Intel Xeon 6 CPUs, and SentinelAI, a control layer for enterprise AI security and governance. On October 1, 2026 it announced a partnership with Muxunav to build models for identity verification, fraud prevention and voice assistance in unattended payment devices. A published Telefónica testimonial reports up to 75% lower energy use than uncompressed models on its network.
In July 2026 the company announced a new funding round, described in its release as a Series C, built around two theses: AI running on devices at the edge, and efficient sovereign AI at scale. The Spanish state investor SETT is listed among its backers. Beyond language models, the company runs Singularity, a quantum-inspired optimization platform.
The other side is worth weighing. The homepage publishes no flat API price, so the cost savings are vendor figures to test on your own workload, and a compressed model can trail its base model on some tasks, since even the company says within 5%, not equal. Teams that only call hosted frontier APIs from OpenAI, Google or Mistral and never self-deploy gain little from compression, and open models from DeepSeek can run uncompressed when hardware is not the constraint. Open weights are usually distributed through Hugging Face, so check each model card for its licence. To find a compressed model that fits your hardware and budget, Smart Match compares options across the directory.
Products
- CompactifAI (Model compression platform and API): https://multiversecomputing.com/compactifai/deployment
- Comet 60B (Open multimodal model (compressed Nemotron 3.5 Super VL)): https://multiversecomputing.com/resources/multiverse-computing-launches-comet-60b-in-collaboration-with-nvidia
- Pulsar 16B (Open model (compressed Nemotron 3 Nano)): https://hokai.io/hub/models/pulsar-16b
- SentinelAI (Enterprise AI security and governance layer): https://multiversecomputing.com/sentinelai
- Singularity (Quantum-inspired optimization platform): https://multiversecomputing.com
Multiverse Computing Models on Hokai
Links
Frequently Asked Questions
What does CompactifAI cost to use?
The company publishes no flat price on its homepage. New CompactifAI API users get the first month free, and the company says its models cut inference costs by 50-80% against uncompressed ones. Treat that savings figure as a vendor claim and test it on your own workload before committing.
Can I try Multiverse Computing models without paying?
Partly. The API offer includes a free first month for new users, and the company describes Comet 60B as open source and available to the community. Check each model card for its licence and terms before any commercial use.
How is tensor-network compression different from quantization?
Quantization lowers the numerical precision of weights. CompactifAI restructures the internal weight matrices into compressed tensor representations, which the company says is why size drops by 80-95%. It then also quantizes, so the models ship in BF16, FP8 and NVFP4.
How much funding has Multiverse Computing raised?
On July 27, 2026 it announced a Series C of $570M (EUR 500M) at a $1.7B pre-money valuation, co-led by Forgepoint Capital International, BNPP SIVF and Bullhound Capital. The company says total funding is expected to reach about $800M including earlier rounds, and the round may stay open to more strategic investors.
How do I get started with CompactifAI?
Sign up for the CompactifAI API through the company dashboard, or run one of its open models such as Pulsar 16B or Comet 60B on your own hardware. Comet ships in BF16, FP8 and NVFP4, so you can match the precision to your GPU and latency targets.