How they use Nebius AI Cloud

Unum trains compact multimodal models on Nebius infrastructure. Nebius reports four jointly trained and open-sourced models: UForm-Gen2, UForm-Gen2-dpo and two UForm-VL-English variants. UForm-Gen2 hybrid training took about one day on eight H100 GPUs, with additional preparation for its individual components. Unum says UForm-Gen-2 uses a 0.5-billion-parameter language model derived from Alibaba’s Qwen1.5, a larger multimodal dataset and an improved vision tower. UForm-Gen2-dpo uses Direct Preference Optimization on VLFeedback and LLaVA-Human-Preference-10K for image captioning and visual question answering. Unum divides experiments between Nebius, other clouds and its own infrastructure.

Unum also used Nebius AI Cloud for StringZilla’s GPU optimization. The year-long port culminated in v4 and added hardware-specific kernels for biological sequence alignment scoring and fingerprinting. Fingerprinting supports retrieval, while pairwise scoring supports reranking. Nebius reports a 300,000-fold speedup over traditional tools for the StringZilla work. Ash Vardanian credits Nebius GPU capacity with supporting Unum’s model pre-training and performance engineering. Unum Cloud separately credits Nebius with supporting USearch v2.25, which added Float8 support and more than 1,000 kernels across processor architectures. Vardanian also benchmarked AffineGaps on a Nebius node with Intel Sapphire Rapids CPUs and NVIDIA H100 GPUs. His Mojo implementation extends the toolkit from proteins to RNA folding and cofolding, with cubic and sixth-degree polynomial operations.

Vardanian reports that the first 10 billion SmallField molecular shapes, representing about 460 billion atomic coordinates, are available on Nebius, AWS Open Data and Hugging Face. Unum’s USearchMolecules repository documents a Nebius Object Storage mirror containing Parquet shards, prebuilt .usearch indexes and .smi files. The three mirrors use identical paths and hold byte-for-byte identical files.