Local Office brings fully local doc AI to Windows ARM64: Qwen3 4B at ~80 tok/s on NPU
HowDevelop · x · 2026-11-03
Developer mfranzon's Local Office — a fully local AI for office documents — now runs on Windows ARM64 (Snapdragon) via Qualcomm's GenieX runtime, after Mac and Ubuntu. On an RTX 3060 (6 GB) laptop, Qwen3.5 4B answers at 80 tok/s with OCR and search on CPU; scans, PDFs and invoices get cited answers with nothing leaving the machine. Screenshots show per-task model assignment (Qwen3 4B assistant, LightOnOCR-2 1B for OCR, Nomic Embed v1.5 for search, Qwen3 VL 4B for charts) all on NPU, with models tested before switching.
More from Infra
- RTX 5090 with 64GB and 96GB VRAM 'confirmed' by US GPU repair tech — MundanePercentage674 · 2026-11-03
- Samsung's second Taylor fab fully booked for AMD and NVIDIA HBM base dies before construction even begins — zephyr_z9 · 2026-11-03
- QLink insight: one model's store is already a valid memory write for another — zeeshanp_ · 2026-11-03
- All six collectives on 4 QwenGPUs match real NCCL bit-for-bit on GB200s — zeeshanp_ · 2026-11-03
- QLink compiles a GPU into Qwen weights and runs CUDA kernels via token decoding — zeeshanp_ · 2026-11-03
- UK's Isambard-AI offers startups up to 1M free GPU hours, Foundry tier 20k-200k, deadline Friday — bodonoghue85 · 2026-11-03