Ling-3.2-flash: Ant's 560B Multimodal Model With a 1M Context Window for $0.30

Ant Group has a new flagship model, and it arrived with almost no warning outside China. Ling-3.2-flash launched on November 5, 2026, as the latest generation of Ant's Ling series: a 560-billion-parameter mixture-of-experts model that takes text, images, and video as input, holds up to one million tokens of context, and costs $0.30 per million input tokens through Ant's Bailian platform API.

That combination is the story: frontier-scale capacity, native multimodality, and a price that sits well below most Western frontier models. The catch is that nearly everything impressive about it comes from Ant's own numbers, and the model is API-only, with no open weights announced. This article lays out what Ling-3.2-flash is, what it costs in dollars, pounds, and Canadian dollars, how it compares with the models we have already covered, and what Ant has not proved yet.

What Ling-3.2-flash is

Ling is Ant Group's general-purpose model series, built on a mixture-of-experts (MoE) architecture. In an MoE model, each token is routed to only a small set of specialist sub-networks, so the model keeps a large total capacity while activating a fraction of it on every step. Ling-3.2-flash is the newest and largest multimodal entry in that line.

The 560B total with 25B activated is the same headline spec as its predecessor, Ling-3.1-flash. The difference is what the model can take in: Ling-3.2-flash fuses vision and language at the architecture level, so images and video are a native input to reasoning rather than something bolted on through an external vision module. That is the whole point of this release, and it is also the part most likely to be misunderstood, so the next section takes it apart.

One number deserves a note before going further. The one-million-token context window is the model's ceiling according to Ant's own documentation. Whether the hosted API exposes the full million tokens from day one is a separate question, and Ant has not answered it in the English-language material available at publication. Its predecessor followed exactly that pattern: a 256K-token cap during the free trial, with the full window promised once billing started. Treat 1M as the ceiling, not a day-one guarantee.

What "native multimodal" means here

Ant says Ling-3.2-flash targets four capability areas, and each one is a different kind of work:

• Complex visual understanding. Deep reasoning over images and video, not just captioning. Ant's benchmark claims sit mostly here.
• Multimodal long-horizon Agent. Multi-step tasks where the model holds a goal across many rounds of tool use, making decisions from combined visual and text information.
• 3D understanding and interaction. Parsing three-dimensional scenes and spatial relationships, aimed at 3D-related interaction work.
• Medical multimodal. Understanding medical images and interpreting reports, to assist professional medical analysis.

The medical item needs a plain caveat. Nothing in the launch material shows clinical validation, regulatory clearance, or independent testing of the medical claims. It is a capability target Ant lists for the model, not a certified medical tool, and it should not be used as one. This is general information, not medical advice.

For most readers, the practical read is simpler. A model that can watch a video, read a document, and keep working across a long task without losing the thread is built for agent workflows: long documents, video analysis, GUI automation, and front-end code built from a screenshot. Ant's own documentation lists exactly those uses, and its DesignArena result (below) is for image-to-frontend work.

What it costs

Ling-3.2-flash's API is priced at $0.30 per million input tokens and $0.90 per million output tokens, through the Ant Bailian platform. Access is free for the first two weeks after launch, which runs to roughly November 19, 2026, calculated from the November 5 launch date. The rates come from IT之家's report of the Bailian listing; Ant's own English-language pricing page was not available to verify against at publication, so treat the figures as reported rates rather than a number pulled from a vendor price page.

In other currencies, at November 6, 2026 market rates (1 USD = £0.747 = C$1.41, verified at publication), the rates work out as follows:

What does a million tokens actually buy? As a rough rule of thumb, a million tokens is around 750,000 English words, so $0.30 covers roughly a small library of input text, and $0.90 covers around 750,000 words of generated output. The ratio matters more than the absolute numbers: output costs three times input, which is the standard shape for API pricing, and it means long generations, not long inputs, are where the bill grows.

There is no subscription tier to buy and no per-seat price. As far as the launch material shows, this is a pay-as-you-go API with a free trial window, and the $0.30/$0.90 rates are the list rates, not an introductory price that rises later. If Ant publishes a different rate card, the free window is the part to watch: it ends around November 19, and after that the list rates apply.

How the price compares

Per-token prices are not like-for-like across models, because the models do different work. Still, the table shows where Ling-3.2-flash sits among the models we have covered recently:

Two things stand out. First, Ling-3.2-flash is priced identically to its text-only predecessor, so Ant is charging nothing extra for the multimodal upgrade. Second, DeepSeek's flash tier undercuts it on input at off-peak hours, but DeepSeek doubles its rates at peak times while Ling-3.2-flash has no peak/off-peak split in its published rates. The lowest headline rate is not always the lowest bill; it depends on when you run and how much output you generate.

The DesignArena leaderboard adds a useful wrinkle. Its notes describe Ling-3.2-flash's rate as the lowest among models of the same performance tier, which is a narrower claim than "cheapest model" and a more honest one. A small model that costs less per token but needs three attempts to get the right answer is not cheaper in practice.

The benchmarks, and why to read them carefully

In Ant's published evaluation, Ling-3.2-flash beats Anthropic's Claude Opus 5 at its maximum setting on three benchmarks:

It also entered the top ten of the DesignArena image-to-frontend arena with an Elo score of 1238, tied with models from Meta, DeepSeek, and Anthropic. That last result is the most concrete of the four: DesignArena is a public, head-to-head arena, and a top-ten Elo in image-to-frontend work lines up with the model's stated focus on turning screenshots into working pages.

All four numbers still need the same flag. These are Ant's own evaluation results, published by Ant, with no independent replication yet. Beating Opus 5 on three benchmarks Ant chose tells you where Ant thinks its model is strong; it does not tell you how the model performs on the task you actually need. OmniDocBench tests document understanding, ZEROBench tests visual reasoning, and Vision2Web tests building web pages from images. Those are narrow slices, and a model can lead on all three while lagging somewhere Ant did not publish.

The honest way to use these numbers is as a reason to try the model during its free two weeks, not as a verdict. Run your own documents and screenshots through it, and compare the output against whatever you use today.

Ling-3.2-flash vs Ling-3.1-flash: same spec sheet, different model

The most confusing thing about this launch is how similar the two models look on paper. Same 560 billion total parameters, same 25 billion activated per token, same one-million-token context window, same $0.30/$0.90 API price. They are not the same model.

Ling-3.1-flash is a text hybrid-reasoning model built for everyday office work, long software development tasks, and agent workflows. Ant published its weights on Hugging Face under the MIT licence in mid-October, with deployment recipes for SGLang and vLLM and a reference setup of eight H200 GPUs. Anyone with the hardware can run it.

Ling-3.2-flash adds native multimodality to the same scale, and takes it away in another direction: there is no open-weights release announced, so the only way to use it is Ant's hosted API. If your work needs images or video as input, 3.2 is the model. If you need to run a model on your own hardware, 3.1 is still the only one of the two you can download.

The 3.1 history also carries a warning about the 1M context window. Ant announced 3.1 on September 30 with a two-week free trial capped at 256K tokens, and said the full one-million-token window would open once billing started. The weights then arrived in mid-October. A ceiling in the documentation and a window in the product are two different things, and 3.2 buyers should check what the API actually serves before planning around a million tokens.

How to try it

Ling-3.2-flash is available through the Ant Bailian platform, with an online chat page for trying it directly and an API for building on it. The API is compatible with OpenAI and Anthropic protocols, which means existing code written against those APIs can point at Ling-3.2-flash by changing the base URL and model name, rather than being rewritten.

A sensible first run looks like this:

1. Open the Bailian platform and the Ling model page, and confirm the model is listed in your region before building anything on it.
2. Use the online chat page to test the exact task you care about: a long document, a screenshot you need turned into a page, or a video you need summarized.
3. Compare the output against your current model on the same input, side by side, before moving any real workload.
4. Only then wire the API into your code, using the OpenAI-compatible endpoint, and watch the first bills: output tokens cost three times input tokens, so long generations are where the money goes.

One gap to note before you start: Ant has not published English-language sign-up documentation that could be verified at publication, and Bailian is Ant's China-focused cloud platform. International availability and the sign-up flow outside China are open questions, and the free two-week window is the right time to find out whether access works from your location.

What is still unknown

• Independent benchmarks. No third-party evaluation of Ling-3.2-flash has been published yet. Everything so far is Ant's own testing.
• The served context window. One million tokens is the documented ceiling. The window the API actually serves at launch is unconfirmed, and the 3.1 precedent was a lower cap during the free period.
• Open weights. None announced for 3.2, unlike 3.1. There is no indication Ant plans to release them.
• International access. No verified English-language sign-up path or statement on availability outside China.
• Medical validation. The medical multimodal capability is a stated focus area with no published clinical validation.
• Post-trial pricing. The $0.30/$0.90 rates are the list rates, and the free window ends around November 19. No separate renewal price has been published, so there is no intro-versus-renewal gap to watch, but also no promise the list rates hold.

Common mistakes

• Confusing 3.2 with 3.1. Same parameter counts, same context figure, same price, different model. If a download link or a weights file is involved, it is 3.1, not 3.2.
• Treating the 1M window as usable on day one. Check what the API serves before designing a workflow around a million tokens.
• Reading vendor benchmarks as a verdict. Three wins on Ant's chosen tests are a reason to trial the model, not proof it beats Opus 5 at your task.
• Assuming the free window means free forever. Two weeks from November 5, then the meter runs at $0.30/$0.90 per million tokens.
• Assuming open weights are coming. Ant open-sourced 3.1 and has said nothing about 3.2 weights. Plan as if the API is the only access.
• Comparing headline rates without counting output. Output costs three times input, and a verbose model can erase a cheap input rate. Measure your own usage.

Frequently asked questions

How much does Ling-3.2-flash cost?

$0.30 per million input tokens and $0.90 per million output tokens through the Ant Bailian API, free for the first two weeks after the November 5 launch. That is roughly £0.22 and £0.67, or C$0.42 and C$1.27, at November 6, 2026 market rates.

Is Ling-3.2-flash open source?

No. It is available through Ant's hosted API only, and Ant has announced no open-weights release. Its predecessor, Ling-3.1-flash, is open source under the MIT licence on Hugging Face.

How is Ling-3.2-flash different from Ling-3.1-flash?

Both are 560-billion-parameter MoE models with 25 billion parameters activated per token, a one-million-token context window, and the same $0.30/$0.90 API price. Ling-3.2-flash adds native multimodal input (text, images, and video), while Ling-3.1-flash is a text-only hybrid-reasoning model with open weights.

What is the context window?

Up to one million tokens, according to Ant's documentation. That is the model's ceiling; the window the hosted API serves at launch is unconfirmed, and Ling-3.1-flash shipped with a 256K cap during its free trial.

Can I use it with my existing OpenAI-compatible code?

Yes. Ant says the API is compatible with OpenAI and Anthropic protocols, so existing code can switch the base URL and model name. International availability and the sign-up flow are still unconfirmed in English-language sources.

Does it really beat Claude Opus 5?

On three benchmarks Ant selected, yes, according to Ant's own published evaluation: OmniDocBench 1.5, ZEROBench (Sub), and Vision2Web. No independent results exist yet, so treat the claim as Ant's, not as a settled verdict.

Conclusion

Ling-3.2-flash is a serious model at an aggressive price: 560 billion parameters, native image and video understanding, a million-token context ceiling, and a $0.30 per-million-token input rate, free to try until around November 19. The open questions are real, though: vendor-only benchmarks, an unconfirmed served context window, no open weights, and unclear international access. The free window is the honest way to settle them. Try it on your own documents and screenshots, measure the output, and decide from your own results rather than Ant's scoreboard.

For price context, read our DeepSeek V4.1 Flash 1017 pricing breakdown and our Inkling 2 Small launch coverage to see how the rest of this month's model releases are priced.

Sources

• Ant Ling official model documentation — developer.ant-ling.com/en/docs/models/ling/
• TechNode, November 6, 2026 — Ant launches Ling-3.2-flash multimodal model with million-token context — technode.com
• Today for AI (IT之家), November 6, 2026 — Ant Group Releases Ling 3.2 Flash Multimodal Model, Topping Claude Opus 5 on Three Benchmarks — todayforai.com
• Hugging Face — inclusionAI/Ling-3.1-flash model card — huggingface.co/inclusionAI/Ling-3.1-flash
• AI Affairs, October 15, 2026 — Ant Group publishes Ling-3.1-flash weights under an MIT licence — aiaffairs.com
• Frontier Now — Ling-3.1-flash weights: MIT, 256K — frontiernow.dev

Exchange rates: November 6, 2026 market rates, verified at publication (1 USD = £0.747 = C$1.41). Pricing figures for Ling-3.2-flash are as reported by IT之家 from the Ant Bailian listing; Ant's own English-language pricing page was not available to verify against at publication.