berkeleybuild.com · for reporters

Run the AI on your own computer

You've downloaded a project's plan sets from the documents shelf and you want an AI to help you read them. You have three tiers of choice — and for journalism the difference matters, because it decides who else sees your documents and your questions.

Tier 1 — easiest: NotebookLM

Upload the PDFs to NotebookLM, ask questions, get answers with citations back to the page. Ten minutes, no setup. The trade: Google processes your documents and sees what you're asking about. Fine for public plan sets; wrong for anything involving a source.

Tier 2 — independent: an open model on your own laptop

Nothing leaves your machine — no Google, no OpenAI, no Anthropic, no server logs, works on a plane. The open-model world is led by strong Chinese and open-source labs (Qwen, DeepSeek, Kimi/Moonshot, GLM/Z-ai, plus Llama, Gemma, and Mistral), and the tooling has become genuinely easy. Much of this guide follows the working notes of Simon Willison — who also happens to have built Datasette, the tool behind our query pages — and his llm command-line tool.

Setup (Mac, ~15 minutes):

brew install ollama          # the local model runner
ollama pull qwen3:8b          # a strong small model (~5 GB download)
brew install poppler          # gives you pdftotext
pip install llm files-to-prompt
llm install llm-ollama        # connect Simon Willison's llm tool to Ollama

Pick the model your hardware can carry:

8–16 GB RAM (most laptops): qwen3:8b or gemma3:12b — capable readers of extracted text.
32–48 GB (Apple Silicon Pro/Max): qwen3:32b or a DeepSeek-R1 distill — noticeably better reasoning over long documents.
64 GB+: 70B-class models run comfortably; this tier reads like a good research assistant.
Prefer clicking to typing? LM Studio is a free GUI that runs the same models.

The workflow:

# 1. extract the text from a downloaded plan set
pdftotext proj3_plans.pdf plans.txt

# 2. ask questions of it, entirely locally
files-to-prompt plans.txt | llm -m qwen3:8b \
  "List every mention of affordable or below-market units, with sheet numbers."

Plan drawings vs text: plan sets are mostly drawings — pdftotext gets the labels and tables but not the pictures. For the drawings themselves, use a local vision model (ollama pull qwen3-vl and point it at page images), or accept Tier 1 for the graphical sheets.

Tier 3 — heavyweight open models by API

The largest open models (Kimi K2, DeepSeek's full releases, GLM's flagships) outgrow laptops but are served cheaply by many independent providers — you keep model choice and price leverage, though you're trusting a server again. A reasonable middle when a document set is huge and non-sensitive.

The one rule that outranks the tooling: language models are reading assistants, not sources. They misread tables, invent sheet numbers, and state guesses fluently. Every number or quote that will appear in print gets verified against the record itself — that habit is the whole reason this project publishes its database.

Why we wrote this page: a newsroom that depends on any single AI vendor — including the ones we use ourselves — has handed away a piece of its independence. Open models on your own hardware are now good enough for document work, and getting better every quarter.

← Back to the reporters page · Sources and further reading: Simon Willison's weblog and llm documentation — the living reference for local-model practice.