How they use Nebius AI Cloud
distil labs builds specialized models for AI agents on Nebius AI Cloud using synthetic data. Its platform prepares evaluation sets and synthetic training data from production traces, trains and scores small models, and offers deployment behind an OpenAI-compatible endpoint.
For Knowunity’s English quiz generator, distil labs reports post-training NVIDIA Nemotron 3.5 Lightning using NVIDIA NeMo RL on Nebius H100 GPUs. Supervised fine-tuning took about 20 GPU-hours, followed by about 48 GPU-hours of reinforcement learning. The resulting model serves a large portion of Knowunity’s English quiz traffic.
In an evaluation of 114 production requests judged by Claude Opus 4.8, 65% of its quizzes passed all four review criteria, compared with 39% for the incumbent.
Sources
2 sourcesNebius Q3 2026 letter to shareholders
Published 12 Nov 2026 · Nebius Group · PDF
Nebius states that Distil Labs is building specialized models for AI agents on Nebius AI Cloud using synthetic data.
Open original sourceCompany context
distillabs.aidistil labs is a Berlin-based software company operated by Distil Labs GmbH. Its platform helps developers and AI product teams build task-specific small language models for recurring agent tasks. Users provide a task description, examples, or production traces; the platform prepares evaluation and training data, post-trains and scores a model, then supports deployment through an inference endpoint or in a customer’s own environment. The company also offers a CLI and API for running the workflow.
Recent updates
1 updateDistil Labs launches an inference endpoint to capture production traffic
Distil Labs introduced an inference endpoint that captures requests and responses while a customer’s existing LLM continues answering. The captured traffic can then serve as starting data for training a small model.
Distil Labs changelog