Blog
Ling 3.2 FlashInclusion AI Ling200 images one prompt

Ling 3.2 Flash Review: Does 200 Images in One Prompt Work?

Hands-on review of Inclusion AI's Ling 3.2 Flash: an 8-image kebab simulator, invoice extraction, an 82-language naming test, and how it ranks.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Accounting, Finance and Marketing

A woman in a beanie and cropped puffer jacket looks up in amazement at a giant wall of printed photos, with a small laptop and a kebab cone beside her.

How good is Ling 3.2 Flash at images, documents and code?

Ling 3.2 Flash, a model from Inclusion AI, is strong at reading images and documents and turning them into working output, and weaker at open-ended creative variety. In one creator's hands-on tests, it built a working interactive simulator from eight photos, extracted a photographed invoice into clean data with one misread digit, and formatted an 82-language naming task correctly while repeating itself. It is slow when it thinks hard. One caveat: these were single runs by one reviewer, so treat them as impressions, not a lab study.

TL;DR

  • Ling 3.2 Flash accepts up to 200 images in a single request, has a 1 million token context window, takes text and images as input, and outputs text only.
  • In a test with eight photos of making a doner kebab, it produced a working interactive simulator in one HTML file, but it took over an hour (about 1 hour 19 minutes).
  • On a photographed invoice, it extracted the table, checked the math and read handwriting, a stamp and a signature, with one slip: it read the year 2026 as 2016.
  • Turning the thinking mode off made the invoice run take about 3 minutes, which suggests that setting matters a lot for everyday work.
  • On an 82-language naming test, formatting was good but creativity was weak: several languages got identical or near-identical names.
  • Benchmark charts show it leading on document understanding, screenshot-to-code and medical answer correctness, while GPT and Claude still lead on the hardest multimodal exam and some agent-style tasks.

What is Ling 3.2 Flash?

Ling is a family of models from Inclusion AI. The reviewer who tested this release describes it as a hybrid reasoning model, meaning it can think through a problem at length or answer quickly, depending on the setting. It has a 1 million token context window, which is the amount of text and image content it can hold in mind at once. A million tokens is roughly the length of several long books.

The headline claim is the image limit: up to 200 images in one request. You can give it text and images. It only replies in text, which includes code. It does not generate pictures.

The same reviewer notes that Inclusion AI's models are not frontier models, meaning they are not the very top systems from the biggest labs. But their performance has been improving steadily, and this release is described as a step up from the previous Ling version, mostly in vision and coding.

What happened when it got 8 images and one big task?

This was the toughest test. The prompt supplied eight photos covering the steps of making a doner kebab: an empty skewer, marinated meat, spices, the cooked cone on the grill, a carving knife, bread and toppings, sauces, and finished wraps. The task was to look at all eight together, understand each step, and build one interactive simulator in a single self-contained HTML file.

An HTML file is a web page you can open in any browser, so there is nothing to install. The model had to infer the process from pictures and then write working code for it.

The result took more than an hour. At the 45-minute mark it was still thinking, and the finished file landed at roughly 1 hour 19 minutes. The final product worked:

  • Eight bowls on the left that you click to add ingredients.
  • A step-by-step panel on the right: marinate, stack, grill, carve, choose bread, add sides and sauce, serve.
  • A cone shape that looked correct, a grill step where the meat browned and rotated, and a carving step where you drag a knife down the cone.
  • A final screen listing what you had added.

The reviewer called it "not bad at all," and the result does show the model connecting images to a sequence of steps and then turning that into code. The cost is time. For a task like this, an hour of waiting is a real trade-off.

Can it read an invoice photographed on a desk?

This test is closer to what most office workers deal with. The image showed an invoice photographed on a desk, with a table, handwritten notes, a stamp and a signature. The model was asked to extract everything, check the arithmetic, and confirm the balance and the signer.

It returned the header and the full table as clean JSON (a structured text format that spreadsheets and software can read). According to the review:

  • Line amounts, subtotal, tax and total due all checked out.
  • It confirmed the advance payment was exactly 50% and that the handwritten balance was correct.
  • It read the handwritten notes, the red stamp and the cursive signature, and matched the signature to the printed name.
  • It noticed that tax applied to the subtotal only.

The slip: it read the invoice number's year as 2016 instead of 2026. That is a small error, but it is the kind that matters. A wrong digit in an invoice number or date can cause a mismatch in your records, and it would not announce itself.

This run took about 3 minutes, with several tool calls, and the reviewer had turned thinking off. His advice: if you do not want overlong responses, turn thinking off or set it low.

How did it do on the 82-language naming test?

The prompt asked for one neighborhood sports team name in each of 82 languages, each rooted in its own culture, all different from one another, in a strict one-line format.

The format and coverage were good. All languages appeared in order, one name each, with the native script and a romanization (the name written in Latin letters). The reviewer noticed some nice cultural touches.

The weakness was variety, and it broke the "all different" rule:

  • Hindi and Urdu got the same name.
  • Punjabi and Saraiki were nearly identical.
  • Indonesian and Malay differed by one letter.
  • Many names were variations on a generic "neighborhood" pattern.

For rarer languages such as Sanskrit and Sinhala, the reviewer did not think the results were good, and invited native speakers to check. That is a fair point: he is not an expert in every language, so the accuracy of the cultural claims is unverified.

The takeaway is that the model followed structure well but did not hold a constraint (every name different) across a long list.

How does Ling 3.2 Flash compare on benchmarks?

The chart shown in the review compares Ling 3.2 Flash with GPT-6 5.6 Sol, Claude Opus 5 K 3, GLM 5.23 flash and Deep Seek V 4.0 flash, all at their maximum settings. The exact model names and scores were not readable enough in the source for me to list them, so here is the pattern as described:

AreaWhere Ling 3.2 Flash stands
Document understandingTop spot
Screenshot of a website to working codeTop spot
Medical answer correctnessTop spot
Visual mathHolds its own
Hardest multimodal examTrails GPT and Claude
Some agent-style tasksTrails GPT and Claude

The summary is that it shines where vision meets documents and code. It does not win everywhere. As the reviewer put it, benchmarks tell only half the story, which is why the hands-on tests are useful.

What does this mean for you?

If you handle paperwork (accounting, procurement, HR, operations): the invoice test is the most relevant. Models that read photographed or scanned documents and return structured data can save real time. But the 2016 versus 2026 slip is the lesson. Have the model check arithmetic (it did well there), and have a person verify identifiers, dates and invoice numbers before anything goes into your records.

If you build or prototype things: turning a set of images into a working page is a strong demonstration. Screenshot-to-code is one of the areas where it ranked well. Expect long waits when thinking is on.

If you write or brainstorm: do not rely on it for varied, culturally specific creative lists. Check for repeats.

Practical steps:

  1. Start with thinking off or low for routine extraction. Raise it only for hard problems.
  2. Ask for structured output (JSON or a table) so you can check it quickly.
  3. Spot-check the fields that cause the most trouble if wrong: dates, IDs, amounts.
  4. Test with your own documents before trusting any benchmark chart.
  5. For long, multi-image jobs, plan for the model to take a long time and run it when you are not waiting on it.

Could you build your own document-checking tool instead of juggling spreadsheets?

The invoice test hints at a useful pattern: upload a document, pull out the fields, check the totals, and flag anything odd for a person to review. If your team does this by hand into a spreadsheet, you could build a small intake tool for it. With Remy, you describe that tool in plain English: an upload form, a table of extracted invoices, a review queue for items needing a human check, and a login for your team. Remy builds and deploys it as a full-stack app, and it can call AI models for the reading step, with 433 available. You own the resulting app and its code.

Frequently Asked Questions

4 questions

How many images can Ling 3.2 Flash accept?

According to the review, it accepts up to 200 images in a single request, along with text. It has a 1 million token context window and outputs text only, so it cannot create images.

Is Ling 3.2 Flash better than GPT or Claude?

Not across the board. Charts shown in the review put it first on document understanding, screenshot-to-code and medical answer correctness. GPT and Claude still led on the hardest multimodal exam and some agent-style tasks. It is described as a strong, efficient vision model, not a frontier leader.

Why did the kebab simulator take so long?

The model was set to reason at length, and it spent time reading the images, planning and using tools before writing the file. The reviewer saw it still thinking at 45 minutes and finishing at about 1 hour 19 minutes. For the invoice, turning thinking off cut the run to about 3 minutes.

Can I trust it to extract invoices?

It handled the table, arithmetic, handwriting, stamp and signature well, but it misread one year. Use it to do the first pass and check dates, numbers and identifiers yourself, especially where mistakes cost money.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.