How to run deep research locally on your Mac
- Deep research is a multi-step job: plan the question, read several sources, then write up what they actually said.
- The reasoning runs on your Mac; only the page fetches touch the network.
- Answers carry citations, because research you can't check is just confident prose.
- It takes minutes, not seconds — that is what separates it from a single search.
A one-shot answer to a research question is a guess with good grammar. Real research means deciding what to look up, reading more than one source, noticing where they disagree, and showing your work.
What bounds each step, in seconds and tokens
A research loop with no bounds does not fail, it runs all night — and the honest version of “it takes minutes” is a list of the limits that make minutes the worst case rather than the hope. These are the values the shipping code enforces.
| Step | Time limit | Output limit | What it is for |
|---|---|---|---|
| Plan the question | 30 s | 300 tokens | Decide what to look up; a plan that rambles is a plan that has stopped planning |
| Search | 30 s | — | Fetching is the only step that touches the network |
| Summarise a source | 60 s | 250 tokens | Per source, so one slow page cannot hold up the rest |
| Write up | 180 s | 1,200 tokens | The long step, and the one worth waiting for |
| Synthesis pass | 80 s whole, 30 s to first token | 420 tokens | The first-token fuse abandons a pass that has stalled instead of waiting out the budget |
| Iterate (go back for more) | 55 s | 2 more queries, 5 new sources | Bounded so “notice what is missing” cannot become an endless loop |
Two of those rows are worth reading twice. The first-token fuse is separate from the overall budget because the two failures look nothing alike: a pass that is slow still finishes, while a pass that never starts will sit there consuming the entire allowance and produce nothing. And the iterate step is capped at two extra queries and five new sources — the step that makes research deep is exactly the step that, unbounded, makes it interminable.
There is also a floor, not just ceilings: a synthesis shorter than 350 characters is rejected rather than returned. A two-line answer to a research question is the failure mode that looks most like success.
deep_research_v4.py, the bounded-phase orchestrator):
plan 30 s/300 tok · search 30 s · per-source summary
60 s/250 tok · write-up 180 s/1,200 tok · compose
80 s budget with a 30 s first-token fuse, 420 tok cap and a 350-character
validation floor · iterate 55 s, 2 queries, 5 new sources. These are the enforced
values, not a description of intent.What makes it "deep"
A normal answer is one pass: your question in, a paragraph out. Deep research is a loop — decompose the question, retrieve, read, notice what's missing, retrieve again, then synthesise.
The synthesis is where value is added or destroyed. A summary that drops the citation is worse than no summary, because it launders a claim into something that looks settled. Keeping sources attached is the whole discipline.
Running one
- Ask a real question — one that genuinely needs several sources.
- Let it work. You'll see it progress through the steps.
- Read the write-up with its sources and follow the ones that matter.
Give it something with substance. "What's the capital of France" wastes the machinery; "how do these four approaches actually compare" is what it's for.
Being straight about the limits
- It reads the web, so the web's errors are available to it. Citations exist so you can catch that.
- Page fetching needs a connection. The thinking is local; the reading isn't.
- A local model is not a frontier model. On a genuinely hard synthesis you should read the sources yourself — the citations make that possible rather than theoretical.
- It takes minutes. If it returned instantly it didn't do the work.
Common questions
Can a local AI do deep research?
Yes — planning, retrieval across multiple sources, and a written synthesis with citations. The reasoning happens on your Mac; the page fetches use the network.
Does deep research work offline?
The reasoning does, but reading current web pages needs a connection. Offline, the model is limited to what it already knows and what you give it.
Are the sources real?
Citations point at pages actually retrieved during the run, so you can open them and check. That is exactly why they're there.
How long does it take?
Minutes rather than seconds — it is doing several rounds of retrieval and reading, not one.
Try it on your own Mac
One signed, notarized download. No account, no token bill, and it keeps working with the Wi-Fi off.
Free: Nano + Lite. Pro: $249 once · or 4 × $62.25; $124.50 for the first 25 · or 4 × $31.13. macOS 26+. In the US, Klarna or Afterpay at checkout: four payments, two weeks apart. Refund window: 30 days, no questions asked.
Download free (290 MB) Buy ProOn your phone?