In recent weeks, Deriva-1 has improved well beyond what we were aiming for. Deriva-0.7, our latest internal iteration, has shown major gains in long-context understanding, and we are now introducing reinforcement learning as we move closer to the final iteration and release.
Our latest internal build has successfully demonstrated substantial improvements in long-context understanding. These gains are helping the model scale more gracefully toward our one-million-token context window target.
Alongside that work, we have begun introducing RL (Reinforcement Learning) to guide Deriva toward the capabilities we want. One particular goal is a substantial improvement on DeepSWE. We do not train on specific benchmarks; we hope those gains will emerge from broader improvements to a model that is already very strong and impressive in our internal work.
We are using RL rather than ERL (Explanatory Reinforcement Learning) because ERL has not yet been successful with Hikaru. One possibility is that the model is simply too small, but we have not established that as the cause.
We also tested a small iteration of ERL on an upcoming internal model we are calling Cinder. This is a codename: the model is not part of our announced lineup, and we are keeping its actual identity private for now.
Deriva is close to its final iteration and release. We want it to be
as good as it can be without taking too long, while also leaving room
to see which new, powerful models become available before we fully
settle on Qwen3.8-27B as the final base.
Later versions, such as Deriva-2.x, will most likely use a different base. That could eventually include a model we build entirely from scratch, but only time will tell.
For more on the project, see the Deriva-1 model page and Update 2. Our ERL research page explains the method we are testing.