Kora Splits Local AI Across CPU, GPU, NPU and Nearby Machines
Market
Hybrid local-to-cloud inference and heterogeneous edge compute / platform architects placing AI across constrained user, ground and eventual flight-class systems.
Trend
This is a vendor prototype and 2027 product plan, not a production deployment. Kora reports Gemma 4 26B at 25 decode tokens/s and 550 prefill tokens/s with a 32K context on Snapdragon X2 Elite's Hexagon NPU, and projects 70% of managed model calls will run locally; the exact chip, RAM configuration, pricing and independent benchmark are missing.
Tech Highlight
The Linux OS uses a hybrid inference layer to place models across CPU, GPU, NPU and LAN peers, then falls back to cloud models. Its current Qualcomm path uses GGUF plus llama.cpp's Hexagon backend and mainline FastRPC, while planned AMD and Intel paths use different runtimes and reserve their NPUs mainly for embeddings or speech—evidence that portable abstractions still terminate in vendor-specific hardware/software partitions.
6-Month Outlook
Watch for reproducible power, memory and latency measurements on shipping X2 Elite PCs, a documented cloud-fallback policy, and working AMD/Intel backends. Comparable results across vendors would validate a transferable edge control plane; a Qualcomm-only launch or opaque placement decisions would expose lock-in behind the hybrid claim.