SynthLabs
Reasoning-model post-training research · Model developers
How they use Nebius
SynthLabs develops AI reasoning models and post-training systems. It used Nebius’s TractoAI platform to validate nearly 650,000 problems for its Big Math reasoning dataset. Researchers wrote notebook-based map operations while TractoAI scheduled and scaled offline LLM inference. The technical story describes a research partnership and plans for larger training runs. It does not establish deployment of the proposed 128-H100 training configuration.
Nebius also states that SynthLabs trains AI reasoning models using TractoAI, described as a serverless platform on Nebius AI Cloud. The stated research goal is to enable reasoning models to move beyond pattern matching through search and exploration strategies. A Nebius AI Cloud meetup listing identifies SynthLabs as a customer using Nebius AI infrastructure and as a partner.
The authors of Just Enough Thinking separately thank Synthlabs and Nebius for providing compute supporting their Adaptive Length Penalty research.
Sources
2 sourcesNebius co-founder describes SynthLabs collaboration
10 Apr 2025 · Roman Chernin · Web page
Confirms Nebius’s work with SynthLabs.
Open original sourceCompany context
synthlabs.aiSynthLabs develops post-training methods, synthetic datasets and software for improving large language models’ reasoning, alignment and safety. Its work serves AI researchers, developers and organizations training or adapting foundation models. Research releases include Meta Chain-of-Thought, Generative Reward Models, the Big-Math dataset and the PERSONA evaluation benchmark. With EleutherAI, it has added preference learning and reward-model training capabilities to GPT-NeoX. It also offers research partnerships and early access to a managed training platform, alongside tools for evaluation, process reward modeling and customized training workflows.
SynthLabs — AI Post-Training ResearchDocumented financing · private company