Deriva-1.

Overview

Deriva-1 has not shipped. Our latest published internal iteration is Deriva-0.7, described in the September 26 update. It shows substantial internal gains in long-context understanding, and we are introducing reinforcement learning while working toward the final release.

Our current work uses Qwen3.8-27B, selected after evaluating a range of candidates. We are still evaluating the final base, including newer models that may become available. Strong as it is, the base model came with real issues we are working through: reasoning modes that don't scale to the task and are, in places, worse than doing less thinking, and heavy censorship on sensitive topics, common to models built under Chinese regulatory guardrails.

Early iterations introduced problems of our own. Deriva-0.1, trained on a small corpus, saw its hallucination rate spike more than 70% over the base model. Deriva-0.2 brought that back down to about 40% and improved personality. Deriva-0.3 improved the base model's censorship and reasoning-mode issues and brought hallucination back down to roughly the base model's own rate, tracing the earlier spike mostly to synthetic training data. Later work broadened the training mix. Deriva-0.4 and 0.5 did not produce the gains we wanted, while early Deriva-0.6 tests showed more encouraging long-context behavior.

We removed the experimental OMEGA reasoning mode from this stage of the project because it amplified the overthinking problem. See the latest Deriva update for current progress, and Update 1 for the earlier iterations.