Deriva-1
An extensive finetune of Qwen3.8-27B, built to fix overthinking and censorship, and the model Kova is built around.
- Series
- Deriva
- Status
- In development
- Released
- Not yet released
- Announced
- Aug 20, 2026
- Parameters
- 27B
- Context window
- 1M tokensTarget
- Reasoning modes
- 5
- Input
- Text, Image, Video
Benchmarks
- MMLU-ProKnowledgeN/A
- GPQA DiamondReasoningN/A
- SWE-bench VerifiedCodingN/A
- DeepSWECodingN/A
- BrowseCompAgentsN/A
- IFEvalInstruction followingN/A
- RULER (128K)Long contextN/A
- Hallucination rateReliability · lower is betterN/A
Results will be published when Deriva-1 launches.
Overview
Deriva-1 is the first model in the Deriva series, our everyday driver for finetuned tasks. It is an extensive finetune of Qwen3.8-27B, built to fix two problems in the base model: overthinking and censorship. It is also the model Kova is built around.
Built on a strong base
Qwen3.8-27B is an unusually strong model for its size, and we chose it knowing it had real issues we would need to fix. Its reasoning modes did not size their thinking to the task, spending thousands of tokens on trivial questions, and it would deflect or fabricate answers on topics it had been trained to avoid. Deriva-1 addresses both, while keeping the raw capability that made the base model worth building on.
We are still reviewing whether Qwen3.8-27B will remain the final base before release. Later versions are likely to use a different one.
Training
Deriva-1 is trained with supervised finetuning over LoRA and QLoRA runs, and we have begun introducing reinforcement learning to guide it toward the capabilities we want. Synthetic data makes up about 70% of the training corpus, drawn from several models rather than one, alongside human-written data and real validation.
Reasoning
Deriva-1 offers five reasoning modes, so you can trade speed for depth depending on the task. Fixing the base model's reasoning-mode problems, where its highest mode overthought simple requests and produced worse results than its lower ones, has been a focus since our earliest builds.
An experimental mode with a much larger thinking budget, OMEGA, has been removed for this stage of development. We plan to revisit it once Deriva has proved itself, most likely around Deriva-1.1 to 1.2.
Long context
Our recent internal builds have brought substantial gains in long-context understanding. Deriva-0.6 began holding the thread through much longer exchanges, and Deriva-0.7, our latest build, improved on that again. These gains are helping the model scale toward our one-million-token context window target.
Input
Deriva-1 accepts text, images, and video as input.
Availability
Deriva-1 is in development and has not been released yet. Follow its progress in the Deriva updates.
Latest updates
All Deriva newsDeriva-1, Update 3: long context and reinforcement learning
Deriva-0.7 improves long-context understanding as we introduce reinforcement learning, test ERL on Cinder, and approach Deriva-1's final iteration.
Kova Update 2: the big leap
Results from Kova's internal beta: browser use, TV control, day-to-day reliability, and a single request that brought several of those pieces together.
Deriva-1, Update 2: rebuilding around the data
Deriva-0.4 and 0.5 exposed overthinking and data problems. We broadened the training mix, removed OMEGA for now, and began Deriva-0.6.
