Deriva-1 in development.

Deriva-1 is an extensive finetune of the new Qwen3.8-27B model. We targeted two problems in the base model specifically: overthinking and censorship. Training used high-quality synthetic data over LoRA and QLoRA runs; internal readings from those runs show a 7-point improvement in browser control over the base model, the capability Deriva-1 needed to work inside Kova.

OMEGA, an experimental reasoning mode

Alongside the finetune itself, we added OMEGA, an experimental reasoning mode that gives Deriva-1 a substantially larger thinking budget than our other modes. In internal testing, OMEGA sometimes improved results by 4x to 9x over what our xHigh mode produced on the same tasks, a large enough gap that we wanted to share it even before benchmarks are finished.

OMEGA also introduced problems later versions intend to fix: overthinking, degraded performance over long tasks, and degraded performance on animation and UI design, where the model would talk itself into believing an output looked “good” instead of asking the user. We are naming these plainly now rather than waiting for someone else to find them.

Why Deriva-1, and why now

Deriva-1 is also our lead priority under the new order of development we set across the company; see Evaluating scope for the full reasoning. In short: Deriva-1 is the model Kova is built around, so it gets our full attention first.

What comes next

Work continues on reducing OMEGA's overthinking and its long-task and UI-design degradation, alongside finishing the benchmark runs we have not yet published. See the Deriva-1 model page for current capabilities and the benchmark template we will fill in once results are ready.