Streaming chat & tools
Tab, Cmd+K, Composer, chat, and agent steps use approved catalog entries when your plan, keys, and provider uptime allow.
Models
Online and offline — one registry, AUTO routing.
ORQVION ships a master model registry (206 online metadata entries, 49 offline GGUF catalog) with routing modes ONLINE, OFFLINE, AUTO, and MANUAL. Online Beta Ready defaults to AUTO (online-first). Offline inference uses ORQVION AI signed bundles and direct GGUF downloads via llama.cpp — not the Ollama application.
Online
Entitlements on the server. Inference on the desktop.
The license server is a control plane — plans, seats, quotas, and feature gates — not a prompt proxy. Chat and images are not relayed through ORQVION; the desktop calls provider APIs with BYOK keys in your keychain when online paths are active.
206 online catalog
Registry metadata for routing and entitlements. Catalog ids describe capability — they are not all hosted endpoints on ORQVION infrastructure.
Honest online limits
Live execution depends on BYOK, provider health, and plan. OpenAI is not a required core provider. Org policy can forbid external models.
AUTO routing
Right model for the right surface and network.
Task classifier plus hardware and connectivity scores pick online vs offline. Manual model selection always wins. Fallback tries compatible alternates before verified local GGUF or ORQVION AI bundles.
Tab completion
Next-edit suggestions prefer a low-latency catalog model — online or offline per your routing mode.
Agent & Composer
Longer context work can target stronger online/BYOK models or verified local GGUF when OFFLINE or AUTO falls back.
Context budget
@-mentions, rules, and index chunks share a visible context window — no silent truncation in the UI.
Offline & BYOK
GGUF on your machine — optional keys for cloud providers.
Offline: ORQVION AI Lite/Standard plus model-manager HTTPS downloads (pause/resume on `.part` files). BYOK: optional bring-your-own provider keys when org policy allows — stored in the OS keychain.
ORQVION AI bundles
Signed offline bundles for air-gapped or offline-first workflows — llama.cpp inference on your hardware.
49 offline GGUF catalog
Direct HuggingFace-style downloads through the model manager. No `ollama serve`, no port 11434. Legacy ids may contain “ollama” in the name only.
Advanced Local
Settings → Advanced Local enables offline paths. The Ready wizard can skip this at first launch; the offline stack ships in the same desktop binary.