~3B · quantized
Sovra FM Edge 1.0
Edge / Pi-class grounded Q&A and short commands under tight memory budgets.
On-device · offline-capable
A private system that answers domain questions, reasons within bounds, and safely executes validated actions — offline and in real time.
Own on-device foundation models for private knowledge and commands — Edge, Core, and Pro sizes mapped to hardware class.
~3B · quantized
Edge / Pi-class grounded Q&A and short commands under tight memory budgets.
On-device · offline-capable
~8B · quantized
Default for desk and Sovra Core — RAG plus validated command layer on your hardware.
On-device · private by default
~14B–32B class · GPU
Heavier domain reasoning on Core / on-prem GPU when context and VRAM allow.
On-device · larger footprint
Sizes are roadmap parameter classes for on-device deployment. We do not invent third-party leaderboard scores here.
Production-oriented layers — not a single-model chat wrapper.
Sovra FM 1.0 (Edge ~3B · Core ~8B · Pro up to ~32B class quantized) via Ollama / llama.cpp. Optimized per hardware target with latency <200ms first token.
Private document ingestion, local vector DB (FAISS/Qdrant), top-k retrieval with reranking for grounded answers.
Structured JSON intents mapped to validated actions — the differentiator that makes Sovra deployable in production.
Prompt templates, orchestration, guardrails, and deterministic fallbacks for reliability.
REST API for OEM integration plus web chat UI. Phase 2 FastAPI backend; Phase 1 demo on shared hosting.
Local latency tracking, error logging, and privacy-preserving usage analytics — no cloud telemetry required.
Offline-capable assistants, IoT intelligence, automotive prototyping.
Team knowledge bases, meeting intelligence, enterprise RAG at scale.
In-vehicle voice, owner manual Q&A, driver support — custom ECU modules.
Map platform capabilities to a certified stack with live pricing.