Orion V3: A New Design. A New Foundation.

Back to research

Orion 3 is here. The biggest change is the one you see every time you open the app: a redesigned space for talking, thinking, writing, and exploring. The interface is darker, the controls feel lighter, and the conversation takes center stage.

There is a new foundation underneath it, too. Orion now uses a Qwen 3.5-based model, replacing the Dolphin 3 model based on Llama that powered the earlier releases. The core promise remains the same: your conversations run privately, on your device.

V2 → V3: see the redesign

V2 brought Orion Agent and AI Roleplay into the app. V3 refines how those experiences look and feel, with a near-black canvas, translucent input controls, and a quieter use of accent color.

These are real screenshots from each version, with different conversations and selected themes. Compare the layout and controls; the blue-to-purple change is a theme choice. Select a screenshot to view it at full size.

Orion Agent

Orion V2 Agent with a blue accent, a solid composer, and a Shakespeare response — A solid composer over a more visible blue background wash.

Orion V3 Agent with a purple accent, a translucent composer, and a formatted C++ response — A translucent composer, a darker canvas, and more emphasis on the answer.

AI Roleplay

Orion V2 Wizard chat with the keyboard open and a blue accent — The earlier character chat, with the keyboard open.

Orion V3 Wizard chat with a purple accent and styled action and dialogue text — The refreshed conversation view, with the keyboard dismissed.

A clearer visual identity

The crescent now has a brighter lavender-purple identity across Orion’s icon, website, sidebar, and widgets. Loading dots share that same color, giving the app a recognizable signature even when you choose a different conversation theme.

From Dolphin to Qwen 3.5

Earlier Orion releases used Dolphin 3, based on Llama. V3 moves to Qwen 3.5; the current app bundles a compact, quantized 4B model, meaning a four-billion-parameter model stored in a form that uses less memory. It runs locally through the app’s existing inference engine.

That gives us a new base for Orion’s conversations and reasoning experiences while keeping personal prompts and responses on-device. Chat, Agent, Roleplay, and Intelligent Notes remain the places where that model becomes something useful.

The next chapter is already underway

V3 brings the new interface and Qwen-based foundation together. For V4, we are working on our own GPT-style foundation models, pretrained from scratch. Read the V4 research note for the direction we are taking.

Explore Orion on the App Store.

“Welcome to the AGI era.”
Greg Brockman · September 3, 2026
Press briefing, reported by Axios