Local models
Mentions of a business
Find what the sources you already read said about a business, set aside items where the name means something else, and list the few that need doing something about.
What it does
bonfyre monitor reads the item archive, finds every item from the last few weeks that names the business, and asks your local model three things about each one: does the name mean this business or something else, what kind of mention is it, and is there anything to do about it. The answer is a short digest with links, plus a JSON report that keeps every judgement.
When several feeds carry the same headline, the digest shows it once with its publishers, and the report keeps each feed's copy. Copies of one story are not counted as separate reports.
Start the local model
The three commands on these pages ask a model that runs on your own computer. Start it once and leave it running; it loads a model only when a question arrives, and only if there is memory for it.
bonfyre-moe serve --port 7000
bonfyre model local
bonfyre model local lists the models in ~/BonfyreModels, how much memory each needs and which one would answer now. Left to itself, the server loads the largest model that fits. To ask for a particular one, add --model with its file name or part of it, such as --model qwen2.5-3b-instruct. While a command runs it holds the model, so the server does not swap or unload it partway through. The hold ends when the command finishes, or ten minutes after a command stops without saying so. A command that settles every line without the model never loads one. Every answer records the exact model file that gave it.
Watch a business
Give the name and one line that says which business you mean. The line is what lets the model tell a store from the ordinary word.
bonfyre monitor --watch Target \
--about "Target Corporation, the US retail chain" \
--days 30 --out target.md
Add --watch again for other names the business goes by. --out report.json writes the full report with a receipt for each item.
Read the digest
The digest starts with Act on: items the model thinks call for a reply, carry a risk such as a lawsuit, a defect or a price worry, or offer an opening. Then come the remaining mentions by kind (news, reviews, sales and deals, events) and last the items set aside because the name meant something else.
On 2026-10-16 the archive held 23 headlines from the last 30 days containing the word Target. Four were about the retailer, all deals and product drops, and 19 used the word in another sense: price targets, inflation targets, targeted attacks. The digest kept the four and one headline about an analyst's price targets, which it should have set aside, and set aside the other 18.
How well it works
Measured on this computer with Qwen2.5 3B Instruct, a 2 GB model:
| Test | Result |
|---|---|
| 23 Target headlines: retailer or not | 22 right |
| 40 Apple headlines: Apple Inc. or not | the fruit, a casino spam comment, two wiki category pages and an anime title set aside |
| The same 40: what needs action | 11 shown under Act on, including all four headlines a person had picked out beforehand (a criticised printer policy, a design defect, a price rise, a stock warning) and a fifth found afterwards, a possible import ban |
| Time | about 5 seconds an item |
Repeating the business description next to each headline made the small model call nearly everything a match (4 of 23 right), so the description goes in the question and the model reads only the item.
What it can and cannot tell you
Coverage is the sources this installation reads, which are mostly national news, so local businesses get few mentions until local sources are added. An item judged about the business is not a verified fact about it, and nothing is sent or posted. The Act on cut was set on the same Apple headlines it is reported on; a separate run on Target flagged one product drop that needed nothing.