Private AI for small businesses, from setups people actually run.

What's new 🎃

Fresh private and local AI news for small businesses, checked daily. Every item links to its source.

Latest stable versions (checked Oct 17, 2026): Ollama v0.40.5, Open WebUI v0.12.3, AnythingLLM v1.18.0

• Oct 17, 2026 New model

  NVIDIA's Nemotron 3.5 Bolt squeezes a 74B model into 24GB

  NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB on Oct 15. It's a 74B-parameter model (9B active per word) packed into about 24GB with 2-bit quantization, built to run everyday agent tasks locally on RTX and DGX Spark hardware. It reads text, images and video, supports up to 256K tokens of context, and ships under the OpenMDW-1.1 license. Ollama v0.40.5 added support.

  Why it matters: If you already own a 24GB card like an RTX 3090 or 4090, this is a new, fully local option worth testing on your own documents before you buy anything bigger.

  Sources: NVIDIA announcement · Model card on Hugging Face · Ollama v0.40.5 release notes

• Oct 17, 2026 Privacy

  Study: popular AI chatbots share conversation details with ad trackers

  Researchers at IMDEA Networks tested nine consumer AI chatbots and found every one contacted at least one third-party advertising or tracking company. Six of nine web apps passed conversation links, auto-generated chat titles, prompts or screenshots to outside firms, sometimes alongside persistent user IDs. Rejecting cookies cut some tracking, but most trackers stayed active.

  Why it matters: Anything you type into a consumer chatbot about clients, health or money may travel further than you think. A model running on your own machine doesn't load ad trackers at all.

  Sources: IMDEA Networks summary · Full paper (PDF)

• Oct 17, 2026 Tool update

  Open WebUI v0.12.3 adds end-to-end encrypted chats

  Open WebUI v0.12.3 (Oct 16) lets members of direct messages and private group channels start a temporary end-to-end encrypted chat. Messages are encrypted in the browser and passed through the server without being saved. It's off by default: an admin turns it on under Admin Settings, and it needs HTTPS in production.

  Why it matters: Handy if your team uses Open WebUI for internal chat and wants extra protection for sensitive conversations. As always, update with docker compose pull and read the release notes first.

  Sources: Open WebUI v0.12.3 release notes

• Oct 17, 2026 Big open models

  Meta's Muse Spark Open and Reflection's Beam: open, but not small-business sized

  Meta released Muse Spark Open on Oct 15 (816B parameters, 42B active, 1M-token context) under its Muse Model License, which allows free commercial use with attribution and limits for large model-as-a-service providers. Meta says it runs locally on two M5 Ultra Mac Studios; otherwise it needs about 700GB of GPU memory. Reflection released Beam on Oct 16 (501B parameters, 23B active) under the Apache 2.0 license, aimed at coding and agents.

  Why it matters: These are big wins for open AI, but they need server-class hardware. The free hosted versions some services offer run in someone else's cloud, so they aren't private in the way a local setup is.

  Sources: Meta: Muse Spark Open · Hugging Face: Welcome Beam

• Oct 17, 2026 Beginner tool

  Llmspeed shows how fast a local model will feel before you buy

  Llmspeed, shared on Hacker News today, lets you pick a GPU, Mac or mini PC and a model and see a typing-speed preview of the replies. It covers hardware from an RTX 3060 to a Mac Studio and labels which numbers are measured from community benchmarks and which are estimates.

  Why it matters: A quick way to sanity-check a hardware purchase before spending money.

  Sources: llmspeed.dev · Hacker News discussion