Local Vision Models in Practice: Automatically Finding the Lowest Price for the Same Item on AliExpress

The AliExpress Price Trap: The Price You See Is Rarely the Final Price

Search for a phone charger on AliExpress and the results page shows $1.09. Click through and you discover that price corresponds to a 25cm USB-C cable in the same listing. The actual charger costs $2.55, and you'll pay extra for shipping at checkout. This isn't an isolated case — according to the author's testing on October 14, 2026, checking three charger listings in a row revealed massive gaps between the search page price and the actual price: the listing priced at $1.09 couldn't buy you a charger at all, and the same held for those priced at $2.28 and $1.15, with real prices of $4.09 and $3.16 respectively.

The common tricks include: search results only showing the cheapest variant in a listing, which is often not the product described in the title (it might be a cable, an adapter, or a different plug type); prices that don't include shipping; some prices being "new customer only" discounts available just on a first purchase; and "sort by price" having virtually no real effect.

To actually find something cheap, you have to manually open 20 tabs, go through each listing one by one, and calculate the item plus shipping yourself. This is the author's routine whenever buying electronic parts every week.

pricehunt: Automating Price Comparison with a Local Vision Model

To solve this, the author built a free Chrome extension called pricehunt. The workflow is straightforward: open an AliExpress product page, select the specific specs you need (wattage, color, plug type, size, etc.), click "Find the same thing cheaper," and the extension mimics how a human would search — opening listings one by one based on images and seller titles, checking all options, and finally returning a list of quotes for the same product, same specs, with shipping included.

The real-world test is convincing: an Essager 33W charger listed at $12.43 on the original page, and the extension took about 2 minutes to find the same item (same model, black, EU plug, free shipping) for $8.42. The author's best manual search result was $8.89.

Each result is clearly labeled: exact match, similar, or different — counterfeits can't slip through; the price is broken down into item price plus shipping; and it flags things like "$8.28 is the US plug price, not the spec you selected," "$1.09 is new buyers only," or "this discount requires 3 different items in your cart."

The whole system is powered by an AI model running on your local machine, with nothing uploaded to the cloud. It judges whether items are truly the same product by comparing product images. Your AliExpress account, cart, and search history all stay local, and searches are free.

Why a Local Model

The author used MiniCPM-V 4.7 for all development and testing — an open-source, free image-understanding model. The extension supports any vision model running through Ollama, or any server compatible with the OpenAI interface. The author also tested other models: Gemma 4 12B correctly identified 45 out of 49 test images, and Gemma 4 31B got 47 right, but both made the mistake of misclassifying plain cables as chargers, whereas MiniCPM-V didn't make that error.

The model is only asked basic yes/no questions: "Same model? Same body? Same port? Same wattage marking?" It turns out wording has a huge impact on accuracy. When asked "Is this the same type of product?" 7 out of 11 different chargers passed; with a stricter phrasing, only 1 out of 11 passed.

There are several technical details worth noting: the extension uses llama.cpp and keeps MoE weights in system memory (--cpu-moe), with VRAM usage of only 4GB on the author's ordinary 8GB GPU. The match percentage comes from the probability value of the "Yes" token (logprobs), and falls back to a simple Yes/No judgment if the server doesn't provide logprobs. The reference product images generate a 256-bit difference hash (after removing the brand banner first), but hash matching is only trusted after the model confirms they're the same class of product.

Live listing testing exposed some bugs: 35 of 59 discount cards had their prices mistakenly parsed from the strikethrough price; the extension added shipping when AliExpress showed free shipping; and the image CDN server returned 403 pages for repeated requests, which were then cached by the browser for a year. These issues were fixed in v1.5 and v1.6 by analyzing developer logs.

Ollama is an open-source framework for running large language models and multimodal models locally. It bundles model weights, an inference engine, and a simple command-line/REST interface together, so users just run ollama run <model-name> to start a model locally and expose an OpenAI-compatible API endpoint. For a browser extension like pricehunt, Ollama's value is that the extension itself doesn't need to embed any inference code or model weights — it just sends standard chat completion requests to http://localhost:11434, and Ollama handles loading the model, managing VRAM and CPU memory, and returning results. This lets users freely switch between different models like MiniCPM-V and Gemma without modifying the extension, and also forward requests to a more powerful machine on the local network, or to any inference server implementing an OpenAI-compatible interface (such as vLLM or LM Studio).

llama.cpp is the lower-level inference engine in this chain, written in C/C++, designed specifically for running quantized models efficiently on consumer hardware — Ollama calls it under the hood. The --cpu-moe mentioned in the article means keeping the massive expert weights of a Mixture-of-Experts (MoE) architecture in system memory, placing only the always-resident parts like attention layers into VRAM, so a multimodal model that would normally require much more VRAM can run on an 8GB GPU, at the cost of slower inference.

logprobs (log probabilities) are the logarithmic form of the probability distribution a language model gives when generating each token. A normal call only returns the finally selected token, but with logprobs enabled, the interface additionally returns candidate tokens and their log probabilities — for example, the model might give -0.02 for the "Yes" token and -4.1 for "No." Taking the exponent of both shows an enormous gap, meaning the model is quite confident the answer is affirmative. pricehunt uses exactly this to soften the binary yes/no judgment into a 0–100% match percentage: the closer the probability is to 1, the more certain the model is that the two images depict the same product. If the server used doesn't provide logprobs for performance or compatibility reasons, the extension falls back to simply reading the "Yes" or "No" string, in which case there are only two levels and no way to reflect the model's confidence, making misjudgments more likely with ambiguous images.

Manual Search Methodology: How the Extension Mimics Human Price Comparison

The extension's search logic comes from the author's six-step manual routine performed every week.

Step one: search using an image of the product in your selected spec. When you enter a product listing, you usually see a collage of all the options first, but an image showing only the spec you need will get more useful results.

Step two: only look at listings for the same product, starting from the cheapest.

Step three: search again using the seller's full title. Many products don't actually have a formal name, and sellers copy each other's titles. Finding one good listing lets you trace your way to others.

Step four: sort the results yourself. Sort by free shipping and paid shipping separately — don't rely on the site's built-in sorting.

Step five: go through every option within a listing, and select only the one you need. 33W is 33W — make sure the plug fits your country.

Step six: repeat the process from the new lowest quote until several consecutive rounds turn up nothing cheaper.

There's also a trick that has nothing to do with AI: sellers don't take their own product photos — everyone uses the same few factory-provided images. The extension remembers the original image of a product and can recognize it when the same image shows up on other products — even when the AI model isn't confident enough due to different angles. However, this has only been tested on one charger series so far, so it's a "promising but unproven" idea.

Difference hash (dHash) is a perceptual hashing algorithm used to quickly determine whether two images are visually similar, rather than comparing them pixel by pixel. The process is to first scale the image down to a very small size (such as 9×8) and convert it to grayscale, then compare the brightness relationships of adjacent pixels row by row: if the left is brighter than the right, record 1, otherwise 0. This produces 8 bits per row and 64 bits across 8 rows, usually combined or extended to get a longer hash string — the 256-bit figure mentioned in the article comes from this. XORing the hashes of two images and counting the number of 1s gives the Hamming distance; the smaller the distance, the more similar they are. It's robust against scaling, mild compression, and brightness changes, but can't distinguish semantics, so pricehunt's process is to first use the vision model to judge "is this the same class of product," and only after the model confirms it does it use the difference hash as supporting evidence — when the same factory image is reused by different sellers, the hashes are nearly identical, which compensates for cases where the model isn't confident due to shooting angle or cropping. The author also removes the brand banner at the top of the image before computing the hash to avoid promotional text interfering with the result.

Who It's For

This tool suits three groups: ordinary consumers who buy a "charger" and end up with a cable; repair shops and electronics hobbyists who need to repeatedly buy the same parts every month; and small sellers who source from AliExpress and resell locally — every dollar saved on sourcing is profit.

For an individual, saving $4 might just be the price of a coffee. But if you need to repeatedly buy 30 parts a month and save $4 on each, that can determine whether a small side business is sustainable.

A single AliExpress listing often contains multiple variations — for example, the same page selling a 25cm short cable, a 1m cable, a 33W charger, a 65W charger, and US and EU plug versions in different combinations, each with its own price and stock. To look cheap in search results, the search engine usually only grabs the lowest-priced variation in that listing to display, and that lowest-priced variation is often an accessory rather than the main product — hence the phenomenon of "searching for a charger but seeing a cable price." The "new customer exclusive price" is a discount the platform sets to attract new users, visible only to accounts that have never placed an order; even if a returning user clicks the same link, they won't see that price, so the same product has different displayed prices for different people. Understanding these two points explains why the author emphasizes comparing prices only after selecting specs, and why sorting by price on the results page is virtually meaningless.

Why Open Innovation Matters Here

AliExpress won't build this kind of tool for you — its "best price" feature serves the platform's own interests. A tool whose sole purpose is to help you find a lower price must be one you can trust. You want it running on your own computer, where you can verify and even replace the model.

Open models bring three advantages. Privacy: the extension browses using your logged-in account, and cookies, cart, and history are never uploaded. Cost: a single search means dozens of image comparisons, so using paid cloud AI would incur a cost for every search; running locally costs only a bit of electricity. Choice: you can plug in any open model and test it yourself, and all test data is labeled with the corresponding model.

The extension loads via Chrome developer mode — download the zip from Releases, unzip it, and select "Load unpacked" in chrome://extensions. The model can be selected in settings between local Ollama or another OpenAI-compatible server.

Author's note: The code and first draft of this article were written collaboratively with Claude in Claude Code. The author provided the ideas, search methods, and key decisions, and verified all results against his own actual purchase records.

One-line Summary

The author built pricehunt, a locally run Chrome extension that uses a vision model to automatically compare AliExpress product images and find the lowest price for the same item and specs including shipping.

Paragraph Summary

The lowest price shown on AliExpress search pages often corresponds to a variant that doesn't match the title and excludes shipping, with the actual landed price potentially several times higher — the author's testing of three charger listings all showed this kind of discrepancy. To address this, the author built the free Chrome extension pricehunt: after selecting specs on a product page, the extension uses the local vision model MiniCPM-V to compare product images one by one, searching through a six-step manual price-comparison process to return a list of exact-match quotes with shipping included, in practice finding the same $12.43 charger for $8.42. The approach emphasizes local execution, no data upload, no cloud fees, freely replaceable models, and reveals the enormous impact of prompt wording on judgment accuracy.