Senon AI
A model family covering language, agents, computer use, and voice, trained in-house so Senon Solutions owns its model repository outright rather than renting capability from a provider whose roadmap it does not set.
The model family behind every Senon Solutions AI system, built in-house so the company owns its own model repository and confidential work never has to reach someone else's.
Core capabilities
- Reason, use tools, and execute agent and software work through Thread-Ripper 1.2
- Understand multiple information types, including native video, through Thread-Ripper 3.5
- Read interfaces through Capybara Vision, pairing a language model with a high-precision OCR pipeline
- Ground actions in pixel space through Platypus where spatial position is what matters
- Transcribe and synthesise speech through Capybara STT and Capybara TTS
- Convene AI councils where several models deliberate before an answer is returned
- Spawn specialised agents under a president agent that synthesises and returns work for re-review
- Serve as the intelligence layer behind VERONICA and other Senon Solutions systems
Why it exists
Two things drove building this rather than serving someone else's weights. Confidential company and client material cannot be sent to a third-party model, which disqualifies the hosted options regardless of how capable they are. Beyond that constraint, Senon Solutions wanted a model repository of its own: models it controls, versions, and improves rather than rents from a provider whose roadmap it does not set.
The family
Thread-Ripper carries language, reasoning, tool use, and agent execution: version 1.2 at 1.1 trillion total parameters with 32 billion active per token, version 3.5 at 1.6 trillion with 49 billion active, both sparse mixture-of-experts across a one-million-token context and both fine-tuned from open-source bases. Thread-Ripper 3.5 adds native video understanding rather than depending on extracted frames or transcripts. Voice runs on Capybara STT at 1.75 billion parameters and Capybara TTS at 3 billion. Computer use is split across two models that see a screen in entirely different ways.
Two ways to see a screen
Capybara Vision is a hybrid: it does not reason about pixel positions at all, instead pairing a language model with a high-precision OCR pipeline to read what is on screen and work out what it means. Platypus operates the way a conventional computer-use model does, understanding pixels directly and grounding its actions in the image itself. Keeping both means an interface can be read as text where text is what matters, and handled spatially where position is what matters, instead of forcing one approach onto tasks it suits poorly.
AI councils
A council convenes several models to deliberate before an answer is returned, and lets a single model spawn specialised agents beneath it. A president agent synthesises what the others found; where one or more of them got something wrong, it sends that work back to be re-reviewed by the agent that produced it rather than silently overruling it. What comes out has been argued over rather than generated once.