Baseten x Parsed Fireside Chat

Amir, CTO of Baseten, sits down with Mudith and Charles, co-founders of Parsed.

Baseten is acquiring Parsed, a team of leading reinforcement learning experts, to advance our strategy as the premier inference company.
Our existing training product was intentionally built at a lower level of abstraction to serve training engineers who know how they want to fine-tune models but need reliable and observable infrastructure rails rather than a “magical” black-box. Customers have consistently told us that they want to eliminate infrastructure headaches, control their models, avoid rate limits, and pay for raw compute instead of per‑token pricing and that’s why we built Baseten Training as it exists today.

However, with the recent advances in fine-tuning and RL, we are convinced that open-source models can now match and potentially outperform closed models for specific, high-value use cases. While closed models may remain ahead on average, the practical gap for targeted tasks can be closed with the right data and training methods. We’ve seen this in the wild: for example, Picnic Health fine-tuned Llama to beat both humans and closed providers at extracting structured data from medical documents.

This is where Parsed comes in.

By integrating Parsed, we are able to combine the right training expertise with the right inference platform to deliver a “magical combo”: tailored, cost-efficient, low-latency models that can be easily deployed at scale. Parsed enables us to serve companies that lack deep internal training talent by partnering hands-on with them to define the problem, understand the data requirements and develop the right techniques to fine-tune the models. Over time, we believe that much of this expertise will be productizable and we will build advanced training workflows increasingly automated and self-serve.