pricehunt 用本機視覺模型搜尋 AliExpress 更便宜的同款商品，並核對規格與運費

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

The problem: the price you see on AliExpress is rarely the price you pay

You're searching for a phone charger on AliExpress and find it for $1.09. You click on it and realize $1.09 is a short cable price for the same listing. The actual price of the charger is $2.55. And you pay for shipping at the checkout.

It happens quite often. On 14 October 2026 I checked out 3 "chargers" from search results:

The usual tricks:

• On search it shows the cheapest in the list. However, often the product is different - cable, adapter, instead of EU - US plug etc.
• the price doesn't include shipping;
• some prices are 'welcome deals' which are only available when you first buy something;
• "sort by price" is not really sorting at all.

So if you want to get it cheaper, you got to open 20 tabs, click on each thing from every listing, calculate the item + shipping costs yourself. That's what I do every week, buying electronics parts from AliExpress.

What I Built

pricehunt is a free Chrome extension that automates all that for you.

Go to the item on AliExpress that you want to buy. Select your specific options (power, color, plug, size etc.) Then click on Find the same thing cheaper.

This extension then diligently searches AliExpress like someone would, by picture, by the seller's title, clicking on the listing one by one and checking all the options. And you finish with a list of offers of the same thing, same option you chose, shipping included.

Genuine search. Found item at $12.43. 2 minutes later the extension finds the same charger (same model, black, EU plug, free shipping) at $8.42. Manually I could find only $8.89 as a best option.

Each result plainly says what it is:

• Exact match, Similar or Different and you can't fake a look-alike
• the price is split into item + shipping;
• like "$8.28 is the US plug price, not the one you selected", "$1.09 is for new buyer only" or "to get this discount, you must have 3 different items in your cart".

This is being powered by a local AI model on your own computer (not the cloud). It compares pictures of the product and determines if they are actually the same product. Your account on AliExpress, cart, searches are all local on your computer. Free search.

Which AI model? I've done the creation and testing using MiniCPM-V 4.7, an open free picture understanding model. It is the best one to use and I've done all the results in this post on it. But it doesn't mean you have to use that. You can hook up any Ollama image understanding model. Or any model that runs through some OpenAI-compatible server. Or any other model. The other ones that I've tested were just slightly less careful (numbers below).

Who it's for:

• anyone that ordered a 'charger' and got a cable;
• repair shops and hobbyists that have to order the same parts every month;
• small resellers from AliExpress buying and selling locally. Every dollar they save buying is a profit.

It's like a coffee to you to save $4 on one charger. But it's the difference between a small side business succeeding or not if you save $4 on every one of 30 parts that you have to reorder each month.

Demo

Before search - your choice (with buttons, 33W, black, EU plug) - this is a picture extension will look for. You can choose a different picture if you prefer.

What's happening in the side panel: how many listings it's looked at so far, which ones it's opened, the current lowest price. If you get a captcha on AliExpress, it will switch that tab to the foreground and wait for you to solve it. No cheating.

Code

helgard-orlm / pricehunt

Find the real cheapest AliExpress offer: the variant you want + shipping, checked by a local open-weight vision model

pricehunt — the same thing, cheaper

A browser extension that finds the real cheapest offer for an AliExpress product: the exact variant you want plus shipping to your country. It's checked by an open-weight vision model running on your own machine.

Open a product page, pick the variant with buttons (power, colour, plug…), press Find the same thing cheaper, and watch the agent work in the side panel.

Built for the DEV Hacktoberfest 2026 Open-Source AI Challenge, week 2 ("Side Hustle"). Started 2026-10-14.

Why

Marketplace search shows the lowest price of any variant, and very often that's not the thing in the title:

…

How to install:

1. Get the zip from Releases. Unzip it.
2. Go to chrome://extensions. Enable Developer mode. Choose Load unpacked and select the folder.
3. Select your AI model in the extension's settings - either your local Ollama on your computer or any OpenAI-compatible server.

How I Built It

The way I manually locate parts. I walked the extension through this process:

1. Search by the photo of the one option you chose. When you go to the list of items you'll usually see a collage of all the options first. And a picture showing only the option you want will have a lot more hits.
2. Look only in the listings that are of the same item, starting with the cheapest.
3. Search again by the seller's exact title. A lot of things don't actually have names, but people will steal the titles from each other. So if you find one good listing you can find the others.
4. Sort the results yourself. Do it for free shipping AND for paid. Don't rely on the sorting on the site.
5. Read every option inside the listing and pick only the one you want. 33W is 33W and make sure you have a plug for your country.
6. Repeat from the new cheapest offer until a few rounds in a row you fail to find anything cheaper.

A trick that has nothing to do with AI. Sellers don't take photos of their own goods - they all use the same few photos from the factory. So the extension will remember which photos your item started with, and when it finds them on other items it'll know, even if the AI model's not confident because the item is at a different angle. (Only tested it on one family of chargers so far so let's call it a promising idea not a proven fact.)

The AI model. Ran MiniCPM-V (v4.7) on my home PC, with a garden-variety 8 GB gfx card. It's an extension that asks it basic, y/n questions about it: "Is it the same model? same body? same ports? same power print?" Turns out it's very sensitive to phrasing. For example, asking "Is it the same type of product?" allowed 7 out of 11 different chargers through. Asking that strict question above only allowed 1 out of 11 chargers through.

What I measured (all with MiniCPM-V 4.7):

What I haven't tested yet: how good it is at spotting truly deceptive listings, where the listing title says "charger" but the image is of a cable. Those 49 images above were all available, with honest titles. So getting two wins against a human doesn't really set any sort of benchmark. It's two wins.

For the technically curious

• Using llama.cpp with mixture of experts weights stored in system RAM (--cpu-moe) so only 4G on the gpu.
• Token probabilities for "Yes" are the match percentages (logprobs). If a server does not provide logprobs then it'll default back to just a Yes/No.
• We take 256bit diffhash of reference product photos (we remove brand banners first). But we only consider this diffhash matched if model also confirms that it's same type of product.
• Real listings testing revealed some bugs: 35/59 discounted cards, the price parser used the crossed-out price. The extension added shipping when AliExpress displays free shipping. And the photo CDN server responds to multiple requests with 403 pages that are cached by browser for 1 year.
• By getting a developer log, it is possible to send the full search trace to some small receiver program on my computer. Most of the bugs that were fixed for version 1.5 and 1.6 were the result of analyzing these traces.

Why Does Open Innovation Matter?

Because the marketplace won't build this for you. The "best price" stuff on AliExpress works for AliExpress. For a tool whose sole purpose is to give you a lower price you have to be able to trust it. You want the tool to run on your computer so you can verify and swap out the model.

Three things were enabled by open models.

• Privacy. This extension uses your login account to browse. No cookies, cart, history etc are ever uploaded off your computer. I'd only use something like this for local model.
• Cost. 1 search = dozens of photo comparisons. If it's a paid cloud AI it would cost money for every search. Here it's just a little bit of electricity.
• Choice. Just add an open model and try it yourself. So Gemma 4 12B has 45/49 on photo test. And Gemma 4 31B has 47/49 on photo test. Both fell into cable only pitfall which MiniCPM-V detected. So all the numbers in this thread have the model in which they were measured.

Co-written with Claude (Anthropic) in Claude Code. Claude wrote most of the code and the first draft of this post. I brought the idea, the search method and the key decisions, and tested everything against my own purchases.

Cover: AI-generated illustration. All screenshots are real.