Tilde Research pretrained language models for a June 2025 sparse-attention study on four Nebius H100 nodes connected by InfiniBand. The experiments used torchtitan with FSDP. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Its Activault repository says most storage testing used Nebius S3, with AWS S3 also tested. Activault stores model activations in S3-compatible object storage and streams them back for interpreter-model training. A March 2025 article credits the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin with helping scale Activault on Nebius.
Nebius also published a Tilde testimonial describing Slurm/Soperator orchestration for model training with real-time streaming of petabyte-scale data. Tilde’s Dhruv Pal described its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.” The testimonial does not identify which Soperator delivery option Tilde used.
Tilde Research pretrained experimental language models using torchtitan with FSDP across four Nebius AI H100 nodes connected by InfiniBand. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Sparse-attention models · Nebius H100 training25 Jun 2025 · Tilde Research · Web pageTilde Research pretrained experimental language models using torchtitan with FSDP across four Nebius AI H100 nodes connected by InfiniBand. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Nebius Tilde · distributed-training testimonial
18 Jun 2025 · Nebius · Web page
Tilde’s Dhruv Pal says Nebius Slurm/Soperator enabled distributed model training with real-time streaming of petabyte-scale data. He describes its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.”
Nebius Tilde · distributed-training testimonial18 Jun 2025 · Nebius · Web pageTilde’s Dhruv Pal says Nebius Slurm/Soperator enabled distributed model training with real-time streaming of petabyte-scale data. He describes its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.”
Activault · Nebius storage testing
27 Sept 2026 · Tilde Research · Web page
Tilde Research reports that most Activault testing used Nebius S3, with AWS S3 also tested.
Activault · Nebius storage testing27 Sept 2026 · Tilde Research · Web pageTilde Research reports that most Activault testing used Nebius S3, with AWS S3 also tested.
Activault · storage scaling account
17 Mar 2025 · Tilde Research · Web page
Tilde Research thanks the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin for assistance scaling Activault on Nebius.
Activault · storage scaling account17 Mar 2025 · Tilde Research · Web pageTilde Research thanks the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin for assistance scaling Activault on Nebius.
Tilde Research is an AI research company developing tools and methods for understanding how neural networks work and improving their design and training. Its work serves machine-learning researchers and engineers studying model behavior, architectures and computational efficiency. Releases include Activault, an open-source engine for collecting and sharing model activations; Sieve, research on steering code generation using sparse autoencoders; and MoMoE, a mixture-of-experts training kernel. More recent work includes attention mechanisms, training optimizers and Popcorn, software that selects validated kernel implementations for particular hardware and inputs.
Model architectures and interpretability research · Model developers
How they use Nebius AI Cloud
Tilde Research pretrained language models for a June 2025 sparse-attention study on four Nebius H100 nodes connected by InfiniBand. The experiments used torchtitan with FSDP. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Its Activault repository says most storage testing used Nebius S3, with AWS S3 also tested. Activault stores model activations in S3-compatible object storage and streams them back for interpreter-model training. A March 2025 article credits the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin with helping scale Activault on Nebius.
Nebius also published a Tilde testimonial describing Slurm/Soperator orchestration for model training with real-time streaming of petabyte-scale data. Tilde’s Dhruv Pal described its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.” The testimonial does not identify which Soperator delivery option Tilde used.
Tilde Research pretrained experimental language models using torchtitan with FSDP across four Nebius AI H100 nodes connected by InfiniBand. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Sparse-attention models · Nebius H100 training25 Jun 2025 · Tilde Research · Web pageTilde Research pretrained experimental language models using torchtitan with FSDP across four Nebius AI H100 nodes connected by InfiniBand. Its 340M-parameter model trained on 15B tokens, and its 1.3B-parameter model trained on 100B tokens.
Nebius Tilde · distributed-training testimonial
18 Jun 2025 · Nebius · Web page
Tilde’s Dhruv Pal says Nebius Slurm/Soperator enabled distributed model training with real-time streaming of petabyte-scale data. He describes its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.”
Nebius Tilde · distributed-training testimonial18 Jun 2025 · Nebius · Web pageTilde’s Dhruv Pal says Nebius Slurm/Soperator enabled distributed model training with real-time streaming of petabyte-scale data. He describes its ease of use, fault tolerance and efficient orchestration as “unique among computing infrastructure solutions.”
Activault · Nebius storage testing
27 Sept 2026 · Tilde Research · Web page
Tilde Research reports that most Activault testing used Nebius S3, with AWS S3 also tested.
Activault · Nebius storage testing27 Sept 2026 · Tilde Research · Web pageTilde Research reports that most Activault testing used Nebius S3, with AWS S3 also tested.
Activault · storage scaling account
17 Mar 2025 · Tilde Research · Web page
Tilde Research thanks the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin for assistance scaling Activault on Nebius.
Activault · storage scaling account17 Mar 2025 · Tilde Research · Web pageTilde Research thanks the Nebius support team and solutions architects Rene Schonfelder and Mikhail Mokrushin for assistance scaling Activault on Nebius.
Tilde Research is an AI research company developing tools and methods for understanding how neural networks work and improving their design and training. Its work serves machine-learning researchers and engineers studying model behavior, architectures and computational efficiency. Releases include Activault, an open-source engine for collecting and sharing model activations; Sieve, research on steering code generation using sparse autoencoders; and MoMoE, a mixture-of-experts training kernel. More recent work includes attention mechanisms, training optimizers and Popcorn, software that selects validated kernel implementations for particular hardware and inputs.
