AI / Deriva Series / Model
Deriva-1.
Our lead-priority model: an extensive finetune of Qwen3.8-27B built to be the dependable, everyday model in the lineup, and the model Kova runs on.
Overview
Deriva-1 hasn't been built yet. What exists today is a line of small internal iterations working toward it, the furthest along being Deriva-0.4, still in progress as of August 26, 2026. Everything below describes where that work stands, not a finished model.
Deriva-1 is being built as a finetune of
Qwen3.8-27B, an unusually strong base model for its
size, chosen after evaluating a wide range of candidates. 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 fixed 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. Deriva-0.4 is now working on the data problem underneath that: stronger guardrails, synthetic data generated across multiple models instead of one, and human-written data with real validation.
Alongside those fixes, we're also exploring OMEGA, an experimental reasoning mode that gives the model a substantially larger thinking budget than the others. It's early and unproven, not yet a confirmed improvement, and carries its own open problems: overthinking, and degraded performance on long tasks and on animation and UI design.
See Deriva-1, Update 1 for the full account of what we've found so far.
Capabilities
What Deriva-1 can do.
Benchmarks
Not published yet.
Benchmark runs are not finished. We would rather leave this table empty than publish numbers we can't stand behind, so every score below is marked N/A until that changes.
| Benchmark | Deriva-1 | Qwen3.8-27B (base) | Comparison |
|---|---|---|---|
| General reasoning | N/A | N/A | N/A |
| Coding | N/A | N/A | N/A |
| Browser control (Kova) | N/A | N/A | N/A |
| Long-context retention | N/A | N/A | N/A |
| Safety / refusal balance | N/A | N/A | N/A |
Deriva-1 has not shipped. The furthest iteration so far is Deriva-0.4, still in internal development, so there is nothing finished to benchmark yet. Every score above stays N/A until a reproducible run exists to publish.
Sential Innovations is an artificial intelligence company and is not affiliated with, sponsored by, or endorsed by the makers of the base model. Any release will follow the base model's original licence, with attribution, and will be provided free and as-is.