Nebius tops rankings in new fine-tuning services benchmark

Fine-tuning is shifting from a research luxury to a core infrastructure requirement for agentic workflows.

• –Nebius emerged as the top choice for iterative workflows, particularly for function-calling tasks and data-centric development.
• –Together AI maintained its lead in raw training speed for dense models but was noted for "rigid" inference formats that cause friction.
• –Tinker offers a significantly lower cost floor ($0.36/1M tokens) but requires manual MLOps management, making it less suitable for teams needing a managed experience.
• –The report highlights a growing trend of using synthetic reasoning traces (distilled from larger models) to train smaller, more efficient specialists.