Freesolo

Freesolo is a platform for post training small language models. Its argument is that a large share of production tasks do not need a frontier model, but smaller models lose quality unless they are tuned, and the effort required to run that tuning keeps teams from trying. Freesolo packages the tuning loop so a model can be specialized for one job and then served at lower latency and cost. The audience is engineering teams shipping narrow features backed by a model. The core offering is Flash, a managed post training package designed to be driven by a coding agent such as Claude Code, Cursor or Codex: you point the agent at the package, it builds a training environment, runs the post training and returns a model behind an endpoint compatible with the OpenAI API. The workspace tracks environments, training runs and live models. The company says it treats kernel engineering as a search problem and optimizes each permutation through an automated research loop to raise training throughput, and the documented workflow combines supervised fine tuning with reinforcement style optimization. Freesolo does not publish a price list or a pricing page. It states that billing is a single quote for a whole run rather than a per token meter or a GPU hour charge, and illustrates this with one example: a run on a model below ten billion parameters over roughly 30 million tokens quoted at $42, against figures of roughly $420 and $450 that it attributes to two alternative services. Treat that as an illustration rather than a published rate card, since the site routes buyers to booking a call.