Multiverse Computing: Efficient AI. Anywhere.

Welcome to the official Multiverse Computing Hugging Face organization. Here you'll find our open compressed models, built with CompactifAI, our optimization technology. Optimized models cost 50-80% less to run, with up to 2x faster inference and accuracy close to the original, so they fit the cloud, on-prem and edge hardware you already have.

Browse the lineup below to find the right model for your use case.

Open-source language models

• Comet 60B A12B: Multimodal model that takes image and text input optimized from NVIDIA's Nemotron-3.5-Super-VL-120B to half its size, maintaining original cappabilities and even improving some key areas. Available in BF16, FP8 and NVFP4.
• HyperNova 60B: Text generation model compressed from OpenAI's gpt-oss-120b to roughly half its size, with lower memory use and latency. Newer dated releases: 2602 and 2605.
• Pulsar 16B: NVIDIA Nemotron-3-Nano-30B-A3B optimized by 50% to 16.15B total parameters (3.1B active per token). Hybrid Mamba2-Transformer with MoE, tool calling, and context up to 1M tokens. Available in BF16, FP8 and NVFP4.

Open-source edge models

• LittleLamb 0.3B: A 290M-parameter bilingual (English and Spanish) model, compressed 50% from Qwen3-0.6B, with thinking and non-thinking modes. Built for on-device and edge deployment. Apache 2.0.

Datasets and research

• llm-refusal-evaluation and safety-for-whom: Datasets for evaluating how LLMs refuse unsafe requests.
• Articles and papers: Notes from our research team on compression, distillation, quantization and safety.

Use our models

• Private deployment: run compressed models in your own cloud, on-prem or on edge devices. Learn more
• Inference API: call compact models through a simple API. Learn more

Website GitHub Contact us