Now in Alpha

AI that actually does the work

OrionNova is a desktop AI assistant powered by Orion - a custom-built model trained for real tasks. Coding. Automation. Research. Not another chatbot.

Join the WaitlistSee How It Works
67%
Benchmark (v1)
Beats GPT-4o
First version
Desktop
Native App
OrionNova
// Real example
You: Refactor this API to use async handlers and add rate limiting
Orion: I'll restructure the endpoints, add asyncio, and implement a token bucket rate limiter.
Reading 4 files... Analyzing dependencies... Writing changes...
Orion: Done. 3 files updated, rate limiter at 100 req/min. Tests passing. Commit?
Features
Built for people who ship
OrionNova isn't a wrapper around someone else's API. It's a complete AI system designed to do real work on your machine.
🖥

Desktop Native

Runs as a real app on your machine. Not a browser tab. Fast, always accessible, integrates with your OS.

🧠

Custom Model

Orion is built from scratch for agentic tasks. It doesn't just answer questions - it takes action, writes code, and manages files.

Real Coding Agent

Reads your codebase, understands context, writes and edits files, runs commands. Not autocomplete. Full autonomy.

🔒

Privacy First

Your code stays on your machine. Desktop-native architecture means your data doesn't live on someone else's server.

🔧

Tool Use

File system access, terminal commands, web browsing, API calls. Orion uses tools like a developer would.

🎯

Personality System

Configure how your AI thinks, responds, and operates. Not a one-size-fits-all chatbot. Your AI, your rules.

The Model
Meet Orion
A custom language model built for real work. Not fine-tuned on vibes. Trained on code, tools, and agentic behavior.

Architecture

Transformer with Grouped Query Attention, RoPE, SwiGLU, and RMSNorm. Designed for efficient inference on consumer hardware with context extension via YaRN.

Training

17.6 billion tokens: code (55%), web text (25%), math (10%), instruction data (10%). Custom 32K vocabulary.

Model Tiers

Spark
Fast
Core
Balanced
Prime
Max power
ORION CORE ARCHITECTURE
Parameters
1.15B
Hidden Dim
2,048
Layers
24
Attn Heads
16
KV Heads
4 (GQA)
FFN Dim
5,504
Vocab
32,000
Context
4,096
TRAINING DATA: 17.6B tokens
Code 55%Text 25%Math 10%Instruct 10%
Benchmarks
Orion v1 beats GPT-4o
46 tests across 14 categories. Orion Core v1 scored 67.0% - outperforming GPT-4o (66.1%) on its very first version.
Claude Sonnet 481.1%
Orion Core v167.0%
GPT-4o66.1%

Strongest: DevOps (83.2%), Complex Scenarios (77.0%), Tool Chaining (75.3%). This is v1. We're just getting started.

Pricing
Simple, transparent pricing
Start free. Scale when you need to.
Spark
Free
  • Basic usage included
  • Orion Spark model
  • Desktop app
  • Community support
Get Started
Starter
$20/mo
  • Extended usage
  • Orion Core model
  • File + terminal access
  • Email support
Join Waitlist
Core
$97/mo
  • Heavy usage
  • Core + Prime models
  • Full tool suite
  • Priority support
  • Custom personality
Join Waitlist
Pro
$297/mo
  • Unlimited-feel usage
  • All models, max context
  • API access
  • Dedicated support
  • Early features
Join Waitlist

Get early access

Be first to try OrionNova. Join the waitlist and we'll notify you when it's ready.