Unbundling Fine-Tuning with MLOps
Kyle Corbitt, CEO of OpenPipe, on the future of fine-tuning LLMs
This category exists because fine-tuning only becomes a real product when someone turns model training into a normal software workflow. The winning move is not selling access to a model, it is giving product teams one place to log prompts, clean data, launch training jobs, compare outputs, and swap a fine-tuned model into production with almost no code changes. That is what unbundling from the model vendors and rebundling with MLOps actually means in practice.
-
OpenPipe starts from live app traffic, not an ML lab. Teams drop in an SDK, capture real requests and responses, turn selected logs into a dataset, relabel weak outputs, run evals, and deploy the result behind an OpenAI compatible endpoint. That makes fine-tuning look like product iteration, not a research project.
-
Predibase is the closest comparable, but it skews more toward ML and platform teams managing many custom models. Its product surface includes supervised fine-tuning, continued pretraining, reinforcement fine-tuning, and multi adapter serving. OpenPipe is more opinionated around app teams that want a faster path from prompt logs to a specialist model.
-
The strategic appeal is buyer shift. Instead of asking a central data science team to build and maintain a pipeline with tools like Unsloth, Axolotl, logging, eval, and hosting, these platforms package the whole loop for the team shipping the feature. That is why funding has flowed into the category, including more than $28M for Predibase and seed backing for Airtrain AI.
Going forward, this market keeps moving up the stack from one off fine-tunes to closed loop model operations. The durable platforms will own the full improvement loop, from production traces to evaluation to retraining to deployment, while model labs and clouds keep pushing down from above with native customization features.