Fast local Qwen
Streaming Qwen 3.8 through MLX Serve on Apple Silicon, with model setup and runtime health owned by Sirius.
0.6 PUBLIC BETA // APPLE SILICON
Sirius gives every project its own constellation of local AI threads—with fast Qwen inference, bounded memory, focused fx execution, and explicit control over tools.
01 // WHAT SIRIUS DOES
Start a clean thread without losing the project. Keep it private, or let it share the useful memory and capabilities of its constellation.
Streaming Qwen 3.8 through MLX Serve on Apple Silicon, with model setup and runtime health owned by Sirius.
Codex-style project threads. Create a new star for a fresh start, then choose whether its memory stays private or shared.
KITKAT and your read-only ChatGPT archive provide cited context only when it is useful—without silently merging every thread.
Turn on workspace tools and longer execution only when the task needs them. Direct chat remains direct.
Research in a Sirius-owned browser profile, separate from your personal browser state and enabled per session.
Turn a message into a Suno track brief, LTX video prompt, ALi render sequence, memory proposal, or reviewed finance draft.
02 // THE REAL PRODUCT
The constellation is not a metaphor pasted onto a chat app. It is how Sirius organizes real project threads, local models, memory scopes, tools, and visible control boundaries.
03 // THE ELASTIC ENGINE
Sirius treats a giant sparse model as addressable components. Routers select the experts required now; bounded storage reads materialize them into MLX; useful components remain warm until memory pressure asks them to leave.

04 // ONE LOCAL CREATIVE SYSTEM
Sirius is the conversational surface. It sends typed, reviewable handoffs to the specialist app that owns the result—never an invisible cross-app mutation.


05 // EXPLICIT CONTROL
fx and macOS Control are separate switches. Enable the agent without handing it your Desktop, then grant Mac control only for the turns that truly need it.
06 // LOCAL BY DESIGN
Sirius is built for people who want a capable AI collaborator without turning every thought into a cloud workflow.
Local inference, deliberate tools, recoverable state. The system stays understandable even as it becomes more powerful.
0.6.1 BETA 1 // AVAILABLE NOW
First launch: open Sirius normally. The notarized app will guide you through local runtime and model setup.
Release notes and SHA-256 →