December 2025
Nebius AI Compute Grant Awarded to Doses AI
Doses AI has received an AI compute grant from Nebius, providing substantial H100 GPU credits for model training. The grant supports the company's research into ternary language models and on-device health AI, enabling large-scale training runs across the full Doses Mortar model lineup.
About the Grant
Nebius — the AI-focused cloud platform that emerged from Yandex's international technology division — provides GPU cloud infrastructure purpose-built for AI workloads. Their AI compute grant programme targets startups and research groups working on foundational AI problems, offering H100 GPU credits that would otherwise represent a significant capital barrier for early-stage companies.
Doses AI was selected for the grant based on its novel approach to model compression. The company's distillation pipeline converts full-precision fp16 models into 1.58-bit ternary architectures — models where every weight is constrained to just three values: {-1, 0, +1}. This extreme quantisation enables the resulting models to run on consumer smartphones, but the training process itself demands substantial GPU compute.
Nebius is an AI-native cloud platform providing GPU infrastructure for AI training and inference. Built on a foundation of large-scale distributed systems expertise, Nebius operates data centres equipped with Nvidia H100 clusters and offers managed ML platforms, storage, and networking optimised for AI workloads.
What the Grant Supports
The Nebius compute grant is being applied to the most compute-intensive stage of Doses AI's model development pipeline:
- Continual pretraining — The core stage of the distillation pipeline where fp16 teacher models are progressively distilled into 1.58-bit ternary student models, requiring hundreds of GPU-hours per model variant
- Full model lineup training — Training ternary models across the complete Mortar family, from the 0.5B parameter variant for ultra-lightweight deployment up to the 12B variant for maximum capability
- Ablation studies — Systematic experiments testing different distillation schedules, learning rates, and architectural choices to optimise the quality-efficiency tradeoff at each model scale
- Domain fine-tuning — Specialising base ternary models for regulated-enterprise tasks including clinical text comprehension and structured document analysis