Together AI becomes data-to-model platform

This acquisition pushes Together from selling raw model access into owning the messy data work that decides whether enterprise AI actually works. Refuel built models and pipelines for turning PDFs, forms, tables, and long text into clean fields that can be used for fine tuning, retrieval, and agent workflows. Folding that into Together lets a customer prepare data and run models in one stack instead of stitching together a separate data vendor.1 2 3 4

• Refuel was not just a small feature add. It came with specialized data cleaning and extraction models, including Refuel LLM-2, and tooling for ingesting and processing large production data pipelines. Together made those models available on its platform for serverless inference and LoRA fine tuning immediately after the deal.1 2 3
• The strategic fit is that Together already sold the compute, inference, and fine tuning layer for open models. Refuel adds the step before training, where teams clean labels, extract attributes, and structure messy source data. That moves Together closer to an all in one workflow for building production AI systems.4 5
• This also shifts Together toward competition with higher level AI tooling companies, not just GPU clouds. CoreWeave and Lambda mainly sell compute capacity, while Together has been moving up the stack toward developer workflows and application infrastructure. Refuel is part of that climb, because better data quality improves model output without needing more GPUs.4 5 6

The next step is for Together to package data prep, fine tuning, and inference as one feedback loop. That makes the platform more valuable as enterprises move from testing generic models to running repeatable workflows on their own documents and records. In this market, the winner is likely the platform that fixes data problems closest to where models are trained and served.1 2 4 5