Kestrel Core

Kestrel Core is Deevo’s ultra-dense Small Language Model (SLM), packaging sub-3B parameter frontier reasoning into a compact footprint tailored for bare-metal on-device execution. Built specifically for edge computing, air-gapped field appliances, developer laptops, and local hardware, Kestrel Core completely bypasses cloud dependencies. It delivers sovereign, low-latency intelligence directly within local memory spaces without leaking an ounce of telemetry to external networks.

The model’s architecture is engineered using native C/C++ runtimes via customized llama.cpp backends, compiled with direct hardware-specific assembly instructions targeting AVX-512, ARM NEON, and Apple Metal silicon. Through deep pruning and structural quantization down to 4-bit and 3-bit precision profiles, Kestrel Core fits within consumer RAM envelopes (requiring less than 3GB of total system memory) while preserving exceptional instruction-following fidelity, syntax parsing, and natural dialogue capabilities.

In operational scenarios, Kestrel Core excels as a high-speed local processing agent. It powers real-time terminal autocompletion, local document red-teaming, confidential record synthesis, and edge sensor analytics at generation speeds exceeding 60 tokens per second on basic hardware. By executing all compute locally, it delivers instantaneous responsiveness immune to network outages, API rate limits, or corporate data privacy breaches.

Kestrel Core represents our foundational answer to the centralization of artificial intelligence. By democratizing high-grade reasoning across compact hardware, it serves as the trusted local node in distributed edge-cloud hybrid topologies, acting as an intelligent privacy shield that cleanses, redacts, and formats local inputs before any communication is routed to larger foundation systems.