CYBERDROID LOCAL INFERENCE ENGINE · CODENAME SIMURG · COMPUTE

Thirty birds. One engine.

“In Attar’s Conference of the Birds, thirty birds cross the world seeking the legendary Simurgh — and find themselves: si murgh, ‘thirty birds.’ Many models, presented as one intelligence behind one API.”

A standalone local inference engine built on llama.cpp: serve quantized open-weight models on your own GPUs behind one HTTP API, so every Cyberdroid agent — and anything you build — can reason locally without embedding an inference stack.

Request a private briefing Read the architecture

Sovereign intelligence, served.

01

Your models, your metal.

GGUF model families served from your hardware — from a single workstation to DGX-class nodes. Full GPU offload or CPU fallback, per-model tuning.

02

Three dialects, one engine.

Native, OpenAI-compatible, and Anthropic-compatible endpoints. Existing SDKs and tools point at local models unchanged.

03

Roles, not model names.

Products ask for a role — agent, explanation, document_translation — and Simurg resolves it to a model family. Swap or upgrade models beneath every product by editing config, not code.

04

Fair by contract.

Per-agent-class API keys with concurrency and queue quotas keep interactive SOC traffic from being starved by background research. Auth is fail-closed.

05

Health you can route on.

A structured health surface — loaded families, queue depth, quota state, degradation — lets orchestrators dispatch local-first and fall back deliberately. Never silently.

06

Speaks your language.

Auto-routing detects the request language, normalizes, classifies to the right role, and answers in the original language.

HOW IT WORKS — ONE CALL, MANY BIRDS

MIMIR NEXUS YOUR TOOLS ONE HTTP API
NATIVE · OPENAI · ANTHROPIC ROLE RESOLUTION
KEYS · QUOTAS · FAIRNESS MODEL FAMILIES
GGUF ON YOUR GPUS · simurgd HEALTH SURFACE
ROUTE-ON, NEVER SILENT

The engine serves; it never decides. Authority lives with Heimdall, approval with humans — Simurg answers the questions it is asked, on your hardware, and nothing more.

Honest facts.

No tokens-per-second theater — we publish behavior, not benchmarks. Sizing is a conversation about your hardware, in the briefing.

DAEMONOne service — simurgd — built on llama.cpp

FLEETConfigurable model families — such as Qwen3, Nemotron (120B-class), and TranslateGemma

LOCALITY100% local; binds to localhost by default

CONSUMERSToday: the SOC Analyst runtime and Agent Central

DIALECTS3 API dialects — native, OpenAI-compatible, Anthropic-compatible

Thirty birds. One engine.

Sovereign intelligence, served.

Honest facts.

Sovereign reasoning, on your GPUs.