Updates

A short record of milestones from Sential. AI development, research, products, and company news, posted as they happen.

Deriva-1, Update 3 Deriva-0.7 brings major gains in long-context understanding. We are introducing reinforcement learning, testing ERL on an internal model codenamed Cinder, and approaching Deriva-1's final iteration. Models Kova Update-2 : The Big Leap The latest internal-beta build brings browser use, reminders, and TV-control testing together in practical day-to-day workflows. Software Internal restructure We are folding the Sential parent-company plan into Sential Innovations, keeping one main company and a clearer focus. Here is what happens to the divisions and Palladium. Company Wanna learn? Why we make free, sourced courses. Updated with the release of The evolution of AI. Company Deriva-1, Update 2: rebuilding around the data Two weeks into Deriva-0.4 and 0.5: an overthinking problem the base model won't give up, a wider and more diverse training mix, the removal of OMEGA reasoning, and an early look at Deriva-0.6. Models Hikaru-1, Update 2: thirty runs of ERL Three months and more than thirty ERL runs: grader poisoning, thin episode data, and the rethink that moved explanations out of the weights and into retries. All internal, nothing released. Models Verplex is cancelled We are stopping work on the Verplex smart glasses and moving that time into our AI models and research. A deliberate scope decision, not a door closed for good. Company Kova Update-1: internal beta begins Kova has entered internal beta: good enough for daily use, still rough around the edges, and tested under tighter restrictions across a wide range of models. Software Deriva-1, Update 1: fixing what the base model got wrong A month into Deriva-1: why we chose SFT over ERL and Qwen3.8-27B as our base, an early hallucination spike, and the reasoning-mode and censorship issues Deriva-0.3 and 0.4 fix. Models Deriva-1 in development Deriva-1, now our lead priority, is an extensive finetune of Qwen3.8-27B targeting overthinking and censorship, with an experimental OMEGA reasoning mode. Models Evaluating scope Too many projects at once slowed everything down. We set a clear order of priority: Kova, then Deriva-1, then Hikaru-1, then PICO-1, and explain why. Company PICO-1 enters development Our live speech-to-speech model enters development, with efficient generation and buffering to smooth brief compute hiccups. Models Hikaru-1, Update 1: reassessing the base model We're pausing to address scope and integrity. After testing 40+ candidate base models, Ministral-3 (8B) replaces IBM Granite 4.1 as the foundation for Hikaru-1.1 through 1.9. Models Building in the open Our site is now public. A plain note on Sential Innovations, an artificial intelligence company, and how to read the honest status labels across the site. Company Hikaru-1 enters development Work begins on our first model, a finetune of IBM Granite 4.1 (8B) trained with Explanatory Reinforcement Learning. The first real test of our central research direction. Models