Ling-3.2-flash: Ant Ling's Native Multimodal Flagship Runs on a Two-Week Clock

The most consequential fact about Ling-3.2-flash is a billing one, and it expires. Ant Ling put the model into its API on 5 November 2026 and made it free for the first fourteen days; when that window closes, "original prices" resume — prices the vendor has not yet published. So every architectural claim below has a short shelf life and one long-lived question attached: what does this cost on day fifteen? Ling-3.2-flash is a 560-billion-parameter mixture-of-experts model with 25 billion parameters active per token, a 1M-token context, and — the actual news — native image, text and video input, which makes it the first Ling flagship that sees rather than only reads. Its predecessor Ling-3.1-flash was text-only. Ling-3.0-flash-VL, the earlier vision model, was a 124B side branch. This is the first time the top of the Ling line and the multimodal capability sit in the same weights.

Ant's own changelog entry (version 2.0.6, dated 5 November) is the release document, and it is one sentence long. The X announcement the same evening adds three benchmark numbers and five demo videos. There is no technical report, no model card, no weights, and no price list. That is a thinner paper trail than Ling-3.1-flash got in September, and it is worth saying plainly before we treat any number as established.

What actually shipped, and what Ant has not said

Four things are documented and checkable today. The parameter count, the active count and the context window appear in two places on Ant's developer site — the Ling model-selection table and the "Ling's Evolution" timeline, which files the 3.2 series under 2026.11. The modality support appears in the API reference: a request to Ling-3.2-flash "natively accepts text, images, and videos," with image_url content blocks for stills and video_url blocks for video. The endpoint shapes are the two Ant already runs everywhere — OpenAI-compatible chat/completions and Anthropic-compatible messages, both at api.ant-ling.com/v1.

Four things are not documented. There is no published per-million-token price for Ling-3.2-flash anywhere on the pricing page — the table lists Ling-3.1-flash, Ling-3.0-flash-VL, Ling-3.0-flash and two Ming image models, and Ling-3.2-flash appears only in the promotional line. There are no open weights: the inclusionAI organisation on Hugging Face carries Ling-3.1-flash, Ling-3.1-flash-1019 and a Ling-3.1-flash base checkpoint, and no 3.2 repository. There is no independent benchmark. And there is no stated output-token cap, which for a model pitched at long deliverable generation is a real omission rather than a footnote.

The weights question has a visible social trail. The only reply worth reading under Ant's announcement is a user asking, in effect, whether it will be open source. Ant has not answered. For a lab whose entire 2026 reputation rests on MIT-licensed 560B checkpoints — Ling-3.1-flash-1019 shipped at 573.8 GB under MIT on 27 October — closing the flagship is a posture change, not a formality.

The three numbers, and how much they are worth

Ant's announcement compares Ling-3.2-flash to Opus 5 running at maximum reasoning effort on three evaluations. All three are vendor-reported, all three come from Ant's own release pipeline, and none has been reproduced outside it. With that label firmly attached:

• OmniDocBench1.5 — 92.21 for Ling-3.2-flash against 90.56 for Opus 5 (max). A document-understanding benchmark; a 1.65-point gap.

• ZEROBench(Sub) — 43.71 against 39.52. This is the widest relative gap of the three, roughly 10.6%, on a 3D-scene benchmark.

• Vision2Web — 85.86 against 80.20. A vision-to-web-implementation evaluation, and the one that maps most directly onto the frontend demos Ant published alongside it.

Read the set carefully and it says something narrower than "beats Opus 5." First, all three benchmarks are visual. Ant picked the three evaluations where a native-multimodal model at 25B active parameters could plausibly beat a frontier closed model, and did not publish a single text-only comparison — no reasoning suite, no coding suite, no agentic terminal task. Second, two of the three (ZEROBench and Vision2Web) are not standard leaderboard instruments with wide third-party participation; a reader cannot easily sanity-check them against other models. Third, "Opus 5 (max)" is a specific configuration of a specific competitor, and the announcement does not state the harness, prompt set, or number of attempts on either side.

That does not make the numbers wrong. It makes them a vendor's claim about a vendor's chosen task set, which is a normal thing for a launch to be and an abnormal thing to repeat as fact. The honest summary is that Ant believes Ling-3.2-flash is the strongest Chinese model available on document, 3D and vision-to-code tasks, and has published exactly the evidence that supports that belief and none of the evidence that would test it.

Where the multimodality is supposed to pay off

Ant names four capability areas, and they are more specific than the usual launch list. The first is complex visual understanding — documents with equations and diagrams, not captioning. The second is "multimodal long-horizon Agent," which in Ant's framing means a model that keeps a visual context alive across many tool calls. The third is 3D understanding and interaction, and the fourth is medical multimodal work.

The demos are the interesting part because they show the intended shape of use. In one, Ling-3.2-flash coordinates specialist agents over shared context to convert Q2 financial data into an analysis report, a six-slide board deck and a highlights video, then revises all three against feedback. In another it drives Blender through UI actions and parameter changes and turns engineering drawings into 3D web models. A third shows it reading headless renders of material spheres and comparing roughness, normal maps and bump detail to refine its own scripts — a visual feedback loop, the same "observe, act, verify, correct" pattern Ant built into Ling-3.0-flash-VL. A fourth reads research PDFs with equations and implements the algorithms in code. A fifth turns meal photos into a weekly food journal with portion estimates and nutrient totals.

Two of those five are worth taking seriously as product claims rather than marketing: the render-inspection loop and the drawing-to-3D conversion both require the model to look at its own output and change course, which is the one capability that separates native multimodal training from bolting a vision encoder onto a text model. Whether it works at Ant's claimed quality is exactly what no independent test has yet established.

The access story is the part with teeth

Ling-3.2-flash is reachable three ways today: the Ling Studio chat product, the OpenAI-compatible chat endpoint, and the Anthropic-compatible messages endpoint. The thinking object is documented as supported across the 3.0, 3.1 and 3.2 series, with enabled as the default and disabled available for latency-sensitive work. Existing SDKs work with a base-URL change.

Then there is the operational reality Ant documents on its own rate-limit page: two queries per second, globally, with 429s above that, and a 90-second timeout on synchronous calls with streaming recommended. For a chat product that is fine. For the long-horizon agent workloads the model is explicitly positioned for — multi-step loops, tool chains, agents that call the model dozens of times per task — two QPS is a hard ceiling that will shape architecture more than any benchmark number will. Anyone planning to run Ling-3.2-flash as an agent backbone should read that page before the pricing page.

On our own side, the honest position: Ling-3.2-flash is not on the OrcaRouter catalogue today. We re-read the live model list on 6 November and it returns 215 identifiers, none of them Ling, inclusionAI, Ring or ERNIE — the only "ling" matches are Kling video models. So nothing here is one call away. What is one call away, if the two-week free window is what you were counting on, are models with published prices you can actually budget against: Qwen3.8-Flash at $0.15 per million input and $0.47 per million output, GLM-5.3-Flash at $0.15 and $0.50, and DeepSeek V4.1 Flash at $0.15 and $0.60. Those are provider list prices passed through at zero markup, so when a vendor cuts a rate it lands on our side the same day rather than after a repricing cycle.

What day fifteen looks like

Ant's pricing page says original prices resume after the promotional period and asks you to "keep an eye on your account balance." That is the entire pricing disclosure for a 560B flagship. Compare the neighbours on the same page: Ling-3.1-flash is listed at ¥2.00 input and ¥6.00 output per million tokens, discounted 75% to ¥0.50 and ¥1.50; Ling-3.0-flash-VL at ¥0.50 and ¥1.50, discounted 72% to ¥0.14 and ¥0.42. If Ling-3.2-flash lands anywhere near its text-only sibling, that is roughly $0.30 per million input and $0.90 per million output at list, before discounts — but that is inference from a neighbouring row, not a published rate, and a multimodal model carries image and video tokens whose pricing Ant has not described at all.

So the reader's decision splits cleanly by timeline. If you have a document-heavy or vision-heavy evaluation to run this week, Ling-3.2-flash is free, capable on paper, and worth the setup cost of a base-URL change. If you are choosing a production backbone for next quarter, you are choosing a model with no published price, no independent score, no weights, a two-request-per-second ceiling and no fallback path — and the responsible move is to run the evaluation now, while it costs nothing, and defer the architecture decision until the price list and the first third-party scores land.

The thing to watch is not the leaderboard. It is whether Ant publishes a rate card and a model card before the free window closes, and whether inclusionAI gets a 3.2 repository the way it got a 3.1 one. A lab that shipped 574 GB of MIT weights three weeks ago and now ships a closed flagship is telling you something about where the margin is. The two-week clock is the deadline for deciding whether you believe it.