LIVE
Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|
Practical recipe · Run locally

A 100% local AI assistant for your confidential documents

Summarise a contract, query a report, draft a reply: on your machine, without a single line leaving it.

Comfortable 40 min

The tools in this recipe

  • Runs the model locally, one command

    FreeLocal or cloud depending on useSecurity: nothing to flag

    See the card
  • Chat interface with document import, wired to Ollama

    FreeData stays localSecurity: one caveat, see the card

    See the card
  • JanApp

    All-in-one alternative if you do not want a terminal (optional)

    FreeData stays localSecurity: nothing to flag

    See the card

Step by step

  1. 1

    Check the machine

    16 GB of RAM minimum for a 7-8B model in Q4, 32 GB for 14B. Hardware Match tells you what fits on your machine before downloading anything.

  2. 2

    Install Ollama and a model

    Install Ollama, then “ollama run” a recent multilingual model of 7 to 14 billion parameters. Test in your language in the terminal before going further.

  3. 3

    Add Open WebUI

    Launch Open WebUI (Docker or pip). It detects Ollama on its own. You now have a ChatGPT-style interface, but at home.

  4. 4

    Import the document

    Drop the PDF into the chat or create a “knowledge base” for several files. First ask a question whose answer you know, to test the reading.

  5. 5

    Work, then verify

    Summary, questions, draft reply. Every claim that matters is checked in the source document: a local 8B model hallucinates more than a frontier cloud model.

Real example

Situation
34-page service contract received from a client, to analyse before signing, on a 32 GB laptop without GPU.
What we asked / set
After importing the PDF into Open WebUI: “List the clauses that bind the provider beyond the service itself (penalties, intellectual property, exclusivity, liability). For each, quote the article number and copy the exact sentence. If you find no clause on a theme, say so.”
What we got
14B model in Q4 (9 GB), answer in 3 minutes. 6 clauses listed with exact article numbers, 1 clause missed (found on rereading), 0 invented clauses. No data left the machine.

Check by hand before using it

  • Every returned quote exists word for word in the PDF (Ctrl+F).
  • The network monitor shows no outbound traffic during the chat (Ollama and Open WebUI do not need any).
  • You reread yourself the sections the model declared “no clause”.

Your data: precautions

  • !Local does not mean infallible: the gain is confidentiality, not quality. Reread.
  • !Open WebUI exposed on the network without a password is a leak: keep it on localhost or enable accounts.
  • !Ollama also offers a cloud mode; check that the chosen model really runs locally (the name without the “cloud” suffix).

Share this verdict

One link, the same shortlist, the same reasons.

Each recipe was actually done. The tools cited are scored and rechecked every week; no placement is sold.

All recipes