Nebius Customer AtlasNot affiliated · Research by Not affiliated · By @anthonybs
Not affiliated with Nebius · Independent research by @anthonybsSee changes
AllTrainLoop

TrainLoop

Researches post-training methods and develops…

How they use Nebius AI Cloud

TrainLoop trained NVIDIA Nemotron 3.5 Super VL with Nolla Health using 32 NVIDIA B200 GPUs supplied by Nebius over approximately one week. The team used GRPO in NVIDIA NeMo RL for visual questions and full simulated patient consultations. NeMo RL colocation allowed every GPU to run rollout inference. TrainLoop reports that changes to output limits and batch size increased rollout throughput from 1,450 to nearly 5,000 tokens per second.

On Nolla’s internal evaluation of 500 visual questions, accuracy increased from 35.0% to 66.8% after visual training. At the 50-step conversational checkpoint, the model completed all 75 held-out simulated consultations, compared with 50 before conversational training. Top-diagnosis accuracy increased from 48% to 64%, while tokens per completed note fell by 47%. Nolla identifies evaluation on real encounters as future work.

Model developersService: Nebius AI CloudTraining & fine-tuningInference

Sources

2 sources

From last place to the frontier in a week

Listed 25 Oct 2026 · TrainLoop · Web page

TrainLoop reports training Nemotron on Nebius-supplied B200 GPUs. It describes NeMo RL rollout inference across the GPUs and throughput improvements following changes to output limits and batch size.

Open original source
From last place to the frontier in a weekListed 25 Oct 2026 · TrainLoop · Web pageTrainLoop reports training Nemotron on Nebius-supplied B200 GPUs. It describes NeMo RL rollout inference across the GPUs and throughput improvements following changes to output limits and batch size.

Training NVIDIA Nemotron for Clinical Care

Published 15 Oct 2026 · Nolla Health · Web page

Nolla Health identifies TrainLoop as its training collaborator and Nebius as the B200 compute provider. It reports internal visual and simulated-consultation results and describes real-encounter evaluation as future work.

Open original source
Training NVIDIA Nemotron for Clinical CarePublished 15 Oct 2026 · Nolla Health · Web pageNolla Health identifies TrainLoop as its training collaborator and Nebius as the B200 compute provider. It reports internal visual and simulated-consultation results and describes real-encounter evaluation as future work.

Company context

trainloop.ai

TrainLoop is a post-training research and product lab that builds specialized AI models for pharmaceutical, biotechnology, logistics and banking enterprises. Its researchers develop training methods, while its deployment team works with domain experts to incorporate their knowledge into proprietary models. Its work includes biological reasoning models for diagnosis and treatment response, agents for financial services and logistics, and multimodal systems for image understanding and document processing. Engagements span defining research objectives with customers through training, deployment and maintaining production performance.

Documented financing · private company

AmountUndisclosedInvestment (stage undisclosed)
· Marble Creek Ventures

Recent updates

1 update

  1. Partnership

    TrainLoop reports Nemotron clinical-model results with Nolla Health

    TrainLoop announced results from training NVIDIA Nemotron 3.5 Super VL with Nolla Health. Visual accuracy increased from 35.0% to 66.8% on 500 internal evaluation questions, while top-diagnosis accuracy rose from 48% to 64% across 75 simulated consultations. These results do not establish performance in patient care.

    TrainLoop
Next, A–ZTranscripta Bio
← Explore all customers

TrainLoop

TrainLoop researches post-training methods and develops specialized models for enterprise partners. · Model developers

How they use Nebius AI Cloud

TrainLoop trained NVIDIA Nemotron 3.5 Super VL with Nolla Health using 32 NVIDIA B200 GPUs supplied by Nebius over approximately one week. The team used GRPO in NVIDIA NeMo RL for visual questions and full simulated patient consultations. NeMo RL colocation allowed every GPU to run rollout inference. TrainLoop reports that changes to output limits and batch size increased rollout throughput from 1,450 to nearly 5,000 tokens per second.

On Nolla’s internal evaluation of 500 visual questions, accuracy increased from 35.0% to 66.8% after visual training. At the 50-step conversational checkpoint, the model completed all 75 held-out simulated consultations, compared with 50 before conversational training. Top-diagnosis accuracy increased from 48% to 64%, while tokens per completed note fell by 47%. Nolla identifies evaluation on real encounters as future work.

Model developersService: Nebius AI CloudTraining & fine-tuningInference

Sources

2 sources

From last place to the frontier in a week

Listed 25 Oct 2026 · TrainLoop · Web page

TrainLoop reports training Nemotron on Nebius-supplied B200 GPUs. It describes NeMo RL rollout inference across the GPUs and throughput improvements following changes to output limits and batch size.

Open original source
From last place to the frontier in a weekListed 25 Oct 2026 · TrainLoop · Web pageTrainLoop reports training Nemotron on Nebius-supplied B200 GPUs. It describes NeMo RL rollout inference across the GPUs and throughput improvements following changes to output limits and batch size.

Training NVIDIA Nemotron for Clinical Care

Published 15 Oct 2026 · Nolla Health · Web page

Nolla Health identifies TrainLoop as its training collaborator and Nebius as the B200 compute provider. It reports internal visual and simulated-consultation results and describes real-encounter evaluation as future work.

Open original source
Training NVIDIA Nemotron for Clinical CarePublished 15 Oct 2026 · Nolla Health · Web pageNolla Health identifies TrainLoop as its training collaborator and Nebius as the B200 compute provider. It reports internal visual and simulated-consultation results and describes real-encounter evaluation as future work.

Company context

trainloop.ai

TrainLoop is a post-training research and product lab that builds specialized AI models for pharmaceutical, biotechnology, logistics and banking enterprises. Its researchers develop training methods, while its deployment team works with domain experts to incorporate their knowledge into proprietary models. Its work includes biological reasoning models for diagnosis and treatment response, agents for financial services and logistics, and multimodal systems for image understanding and document processing. Engagements span defining research objectives with customers through training, deployment and maintaining production performance.

Documented financing · private company

AmountUndisclosedInvestment (stage undisclosed)
· Marble Creek Ventures

Recent updates

1 update

  1. Partnership

    TrainLoop reports Nemotron clinical-model results with Nolla Health

    TrainLoop announced results from training NVIDIA Nemotron 3.5 Super VL with Nolla Health. Visual accuracy increased from 35.0% to 66.8% on 500 internal evaluation questions, while top-diagnosis accuracy rose from 48% to 64% across 75 simulated consultations. These results do not establish performance in patient care.

    TrainLoop

More in Model developers

AdaptionAMI (Advanced Machine Intelligence)Black Forest Labs