YerevaNN trained larger molecular language models on Nebius H100 cloud GPUs using a corpus of 110 million molecules and 40 billion tokens. The corpus enabled prompts about molecular similarity and basic molecular properties. Three models were integrated into a genetic algorithm for molecular optimization. YerevaNN describes training Galactica and Gemma models and periodically fine-tuning the language model with oracle scores to guide molecular optimization. Its paper reports 488 H100 GPU hours for Chemma-2B and 288 for Chemlactica-1.3B. Nebius describes Flash Attention 2, FSDP model sharding and post-backward prefetch, plus text tokenization, data movement and CPU-based molecular docking on its systems. Parallelization reduced 5,000 docking results from approximately 48 hours to 4 hours. An API routed docking results to local servers. Nebius reports an 8% performance improvement over previous methods and processing rates of 180,000 words per second. Its customer story also reports that parallel compressed SCP transfers reached 200 MB/s, compared with 10 MB/s using rsync. YerevaNN’s director describes access to Nebius cloud infrastructure and support. YerevaNN reports that its molecular optimization approach was 2–4 times more sample-efficient on three docking targets than Augmented Memory with Beam Enumeration. It also reports state-of-the-art performance on drug-likeness (QED) optimization, the Practical Molecular Optimization benchmark and several benchmarks involving protein docking simulations.
In GeoCrossBench, YerevaNN studied generalization across satellite sensors and spectral bands by training on one subset of bands and evaluating on other bands or sensors. It reports that DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model, pretrained on satellite imagery, showed stronger transfer-learning results than DINOv3. χViT extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling. YerevaNN says training used H100 GPUs at Yerevan State University and a node donated by Nebius. It released code and model weights.
YerevaNN also credits Nebius with providing compute for hands-on sessions at Armenia LLM Summer School 2025. Participant Silva Vardanyan names YerevaNN and Nebius among the organizers of the completed 2026 school.
Nebius describes YerevaNN’s molecular-model training and CPU-based docking on its systems. Parallel docking reduced 5,000 results from approximately 48 hours to 4 hours. An API routed results to local servers, and parallel compressed SCP transfers reached 200 MB/s compared with 10 MB/s using rsync.
Nebius YerevaNN molecular-model training27 Sept 2026 · Nebius · Web pageNebius describes YerevaNN’s molecular-model training and CPU-based docking on its systems. Parallel docking reduced 5,000 results from approximately 48 hours to 4 hours. An API routed results to local servers, and parallel compressed SCP transfers reached 200 MB/s compared with 10 MB/s using rsync.
Nebius: molecular AI research
27 Sept 2026 · Nebius · Web page
Nebius says YerevaNN used its infrastructure to accelerate text tokenization and molecular docking. It reports an 8% performance improvement over previous methods and processing of 180,000 words per second, with Flash Attention, FSDP and post-backward prefetch supporting preprocessing and training.
Nebius: molecular AI research27 Sept 2026 · Nebius · Web pageNebius says YerevaNN used its infrastructure to accelerate text tokenization and molecular docking. It reports an 8% performance improvement over previous methods and processing of 180,000 words per second, with Flash Attention, FSDP and post-backward prefetch supporting preprocessing and training.
The cost of launching new satellites is rapidly decreasing. In the coming decade, hundreds of new Earth observation satellites will be deployed, many featuring novel sensors and spectral band… | YerevaNN
26 Nov 2025 · www.linkedin.com · Web page
YerevaNN says its satellite-imagery models were trained on H100 GPUs at Yerevan State University and a node donated by Nebius. GeoCrossBench tests training on one subset of satellite bands and evaluation on other bands or sensors. DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling, and outperformed DINOv3 in transfer-learning settings. YerevaNN released code and model weights.
The cost of launching new satellites is rapidly decreasing. In the coming decade, hundreds of new Earth observation satellites will be deployed, many featuring novel sensors and spectral band… | YerevaNN26 Nov 2025 · www.linkedin.com · Web pageYerevaNN says its satellite-imagery models were trained on H100 GPUs at Yerevan State University and a node donated by Nebius. GeoCrossBench tests training on one subset of satellite bands and evaluation on other bands or sensors. DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling, and outperformed DINOv3 in transfer-learning settings. YerevaNN released code and model weights.
YerevaNN releases molecular language models
29 Jul 2024 · YerevaNN · Web page
YerevaNN says its larger molecular language models were trained on H100 cloud GPUs provided by Nebius AI. Development and pretraining of the smaller models used A100 GPUs at Yerevan State University.
YerevaNN releases molecular language models29 Jul 2024 · YerevaNN · Web pageYerevaNN says its larger molecular language models were trained on H100 cloud GPUs provided by Nebius AI. Development and pretraining of the smaller models used A100 GPUs at Yerevan State University.
Small Molecule Optimizationwith Large Language Models
14 Oct 2026 · arxiv.org · Web page
The YerevaNN-affiliated research team reports training Chemma-2B on Nebius.ai cloud for 488 H100 GPU hours and Chemlactica-1.3B for 288 H100 GPU hours.
Small Molecule Optimizationwith Large Language Models14 Oct 2026 · arxiv.org · Web pageThe YerevaNN-affiliated research team reports training Chemma-2B on Nebius.ai cloud for 488 H100 GPU hours and Chemlactica-1.3B for 288 H100 GPU hours.
YerevaNN is an Armenian nonprofit research lab developing machine-learning models for molecular design, robotics and remote sensing. Its molecular research includes Mol-E, which combines the ChemLactica and Chemma language models with evolutionary search, and BARTSmiles for molecular representations. It also develops models for satellite imagery, wireless sensing and autonomous navigation. Its outputs include research papers, models, code and benchmarks for the scientific community, alongside research opportunities and training for university students and early-career computer scientists.
Molecular AI and machine-learning research · Research
How they use Nebius AI Cloud
YerevaNN trained larger molecular language models on Nebius H100 cloud GPUs using a corpus of 110 million molecules and 40 billion tokens. The corpus enabled prompts about molecular similarity and basic molecular properties. Three models were integrated into a genetic algorithm for molecular optimization. YerevaNN describes training Galactica and Gemma models and periodically fine-tuning the language model with oracle scores to guide molecular optimization. Its paper reports 488 H100 GPU hours for Chemma-2B and 288 for Chemlactica-1.3B. Nebius describes Flash Attention 2, FSDP model sharding and post-backward prefetch, plus text tokenization, data movement and CPU-based molecular docking on its systems. Parallelization reduced 5,000 docking results from approximately 48 hours to 4 hours. An API routed docking results to local servers. Nebius reports an 8% performance improvement over previous methods and processing rates of 180,000 words per second. Its customer story also reports that parallel compressed SCP transfers reached 200 MB/s, compared with 10 MB/s using rsync. YerevaNN’s director describes access to Nebius cloud infrastructure and support. YerevaNN reports that its molecular optimization approach was 2–4 times more sample-efficient on three docking targets than Augmented Memory with Beam Enumeration. It also reports state-of-the-art performance on drug-likeness (QED) optimization, the Practical Molecular Optimization benchmark and several benchmarks involving protein docking simulations.
In GeoCrossBench, YerevaNN studied generalization across satellite sensors and spectral bands by training on one subset of bands and evaluating on other bands or sensors. It reports that DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model, pretrained on satellite imagery, showed stronger transfer-learning results than DINOv3. χViT extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling. YerevaNN says training used H100 GPUs at Yerevan State University and a node donated by Nebius. It released code and model weights.
YerevaNN also credits Nebius with providing compute for hands-on sessions at Armenia LLM Summer School 2025. Participant Silva Vardanyan names YerevaNN and Nebius among the organizers of the completed 2026 school.
Nebius describes YerevaNN’s molecular-model training and CPU-based docking on its systems. Parallel docking reduced 5,000 results from approximately 48 hours to 4 hours. An API routed results to local servers, and parallel compressed SCP transfers reached 200 MB/s compared with 10 MB/s using rsync.
Nebius YerevaNN molecular-model training27 Sept 2026 · Nebius · Web pageNebius describes YerevaNN’s molecular-model training and CPU-based docking on its systems. Parallel docking reduced 5,000 results from approximately 48 hours to 4 hours. An API routed results to local servers, and parallel compressed SCP transfers reached 200 MB/s compared with 10 MB/s using rsync.
Nebius: molecular AI research
27 Sept 2026 · Nebius · Web page
Nebius says YerevaNN used its infrastructure to accelerate text tokenization and molecular docking. It reports an 8% performance improvement over previous methods and processing of 180,000 words per second, with Flash Attention, FSDP and post-backward prefetch supporting preprocessing and training.
Nebius: molecular AI research27 Sept 2026 · Nebius · Web pageNebius says YerevaNN used its infrastructure to accelerate text tokenization and molecular docking. It reports an 8% performance improvement over previous methods and processing of 180,000 words per second, with Flash Attention, FSDP and post-backward prefetch supporting preprocessing and training.
The cost of launching new satellites is rapidly decreasing. In the coming decade, hundreds of new Earth observation satellites will be deployed, many featuring novel sensors and spectral band… | YerevaNN
26 Nov 2025 · www.linkedin.com · Web page
YerevaNN says its satellite-imagery models were trained on H100 GPUs at Yerevan State University and a node donated by Nebius. GeoCrossBench tests training on one subset of satellite bands and evaluation on other bands or sensors. DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling, and outperformed DINOv3 in transfer-learning settings. YerevaNN released code and model weights.
The cost of launching new satellites is rapidly decreasing. In the coming decade, hundreds of new Earth observation satellites will be deployed, many featuring novel sensors and spectral band… | YerevaNN26 Nov 2025 · www.linkedin.com · Web pageYerevaNN says its satellite-imagery models were trained on H100 GPUs at Yerevan State University and a node donated by Nebius. GeoCrossBench tests training on one subset of satellite bands and evaluation on other bands or sensors. DINOv3 outperformed remote-sensing-specific foundation models in the ViT-B category. Its χViT model extends ChannelViT with iBOT/DINO-like pretraining and channel subsampling, and outperformed DINOv3 in transfer-learning settings. YerevaNN released code and model weights.
YerevaNN releases molecular language models
29 Jul 2024 · YerevaNN · Web page
YerevaNN says its larger molecular language models were trained on H100 cloud GPUs provided by Nebius AI. Development and pretraining of the smaller models used A100 GPUs at Yerevan State University.
YerevaNN releases molecular language models29 Jul 2024 · YerevaNN · Web pageYerevaNN says its larger molecular language models were trained on H100 cloud GPUs provided by Nebius AI. Development and pretraining of the smaller models used A100 GPUs at Yerevan State University.
Small Molecule Optimizationwith Large Language Models
14 Oct 2026 · arxiv.org · Web page
The YerevaNN-affiliated research team reports training Chemma-2B on Nebius.ai cloud for 488 H100 GPU hours and Chemlactica-1.3B for 288 H100 GPU hours.
Small Molecule Optimizationwith Large Language Models14 Oct 2026 · arxiv.org · Web pageThe YerevaNN-affiliated research team reports training Chemma-2B on Nebius.ai cloud for 488 H100 GPU hours and Chemlactica-1.3B for 288 H100 GPU hours.
YerevaNN is an Armenian nonprofit research lab developing machine-learning models for molecular design, robotics and remote sensing. Its molecular research includes Mol-E, which combines the ChemLactica and Chemma language models with evolutionary search, and BARTSmiles for molecular representations. It also develops models for satellite imagery, wireless sensing and autonomous navigation. Its outputs include research papers, models, code and benchmarks for the scientific community, alongside research opportunities and training for university students and early-career computer scientists.
