Local AI second‑brain in Obsidian
A user built an ADHD‑focused Obsidian second brain running a local Nvidia Nemotron AI for notes, goals and review cycles — fully offline to protect data and latency. (x.com)
NVIDIA describes Nemotron as a family of open models with publicly available weights, training data and recipes intended for building agentic and reasoning-capable systems. (developer.nvidia.com) The Nemotron‑3 Nano variant commonly used for local deployments is published as a ~30B‑parameter family with roughly 3.5B active parameters and a 1,000,000‑token context window, and its model card lists a pretraining cutoff of June 25, 2025. (build.nvidia.com) (huggingface.co) Hugging Face hosts Nemotron checkpoints and community manifests, while runtime projects such as vLLM, Ollama and Docker recipes provide deployment guides and images for serving Nemotron models locally. (huggingface.co) (github.com) (ollama.com) Official and third‑party docs note substantial GPU memory needs for BF16/FP16 runs — tutorials and the DigitalOcean guide recommend A100/H100‑class GPUs or ~60+ GB VRAM for comfortable BF16 inference, while community posts document quantized Ollama or AWQ builds that reduce VRAM enough to run on some high‑end consumer GPUs. (digitalocean.com) (docs.vllm.ai) (tel-zur.net) Obsidian community tooling already targets fully local assistants — the Smart Second Brain plugin thread and several GitHub plugins (e.g., obsidian‑Smart2Brain and NoteGPT derivatives) explicitly advertise offline, vault‑local AI interactions that can be hooked to a local model runtime. (forum.obsidian.md) (github.com) (ssp.sh) Inference stacks like vLLM emphasize throughput and p50 latency reductions via paged attention and continuous batching, and community benchmarks/recipes show those optimizations are the path builders use to keep responsiveness acceptable when the model runs on‑premises rather than via a cloud API. (github.com) (docs.vllm.ai) Multiple tutorials demonstrate wiring Nemotron into a retrieval‑augmented pipeline for grounded Q&A and document summarization—examples include DataCamp and DigitalOcean guides that show how retrieval, local embeddings and a small RAG index are used to connect note collections, scheduled review prompts and goal trackers inside local apps. (datacamp.com) (digitalocean.com)
