Ling 3.2 Flash: A 560B Multimodal Flagship That Watches 30 Seconds at a Time

Ling 3.2 Flash is a 560-billion-parameter mixture-of-experts model from Ant Ling with 25 billion parameters active per token, a 1M-token context window, and — the reason it exists at all — native image and video input rather than a vision encoder bolted onto a text model. It went into Ant's API on 5 November 2026. Nine days later the interesting question is no longer what it is, because the spec sheet is public and short. It is what the spec sheet leaves out, and the first thing it leaves out is that the video understanding everyone is excited about stops at thirty seconds, one clip per request, thirty-two sampled frames.

That is not a footnote. Ant Ling's own multimodal guide — the page that documents how to call Ling 3.2 Flash — sets the video ceiling at a duration of thirty seconds or less, a single video per request, a fixed 2fps sampling rate capped at 32 frames, and a 32 MiB ceiling on the base64 string. The same limits block governs Ling-3.0-flash-VL, so some of it is inherited plumbing rather than a decision made for this model, but it is what a caller gets. If your pitch for a native multimodal model is "long-horizon agent that keeps visual context alive across many tool calls," a thirty-second clip and one video per request is the number that constrains the architecture, not the benchmark table.

What the model is, in numbers you can check

The parameters and the window appear in two places on Ant's developer site: the Ling model-selection table, where Ling 3.2 Flash sits above Ling-3.1-flash with the identical 560B / 25B / 1M row and a different positioning line, and the "Ling's Evolution" timeline, which files the 3.2 series under 2026.11. The positioning line is the actual delta: 3.1 is described as a hybrid reasoning model for office work, long-horizon software development and agent tasks; 3.2 is described as a native multimodal model focused on complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work. Same skeleton, one new sense.

• Total parameters — 560B for Ling 3.2 Flash, matched only by its text-only sibling Ling-3.1-flash at 560B; the vision predecessor Ling-3.0-flash-VL was 124B.

• Active parameters per token — 25B, against 5.5B for Ling-3.0-flash-VL and 25B for Ling-3.1-flash.

• Context window — 1M tokens, four times the 256K that Ling-3.0-flash-VL and Ling-3.0-flash offer.

• Modalities — text, image and video input, text output, with vision and language combined in the architecture itself rather than routed through an external module.

• Endpoint shapes — OpenAI-compatible chat/completions and Anthropic-compatible messages, both at api.ant-ling.com.

The practical consequence of the endpoint shapes is that an existing integration moves with a base-URL change and a model string. Ant ships the same two formats for every Ling model, and the multimodal guide shows image_url content blocks for stills and video_url blocks for video, both base64-only — there is no URL-fetch path documented for either, which is the second constraint a production integration meets before it meets a benchmark.

The price is still unlisted, and the meter has five days on it

This is the part worth checking before you build anything. Ant's pricing page — last updated 5 November 2026, the day of launch — carries rows for Ling-3.1-flash, Ling-3.0-flash-VL, Ling-3.0-flash and two Ming image models. It does not carry a row for Ling 3.2 Flash. The model appears on that page exactly once, in a promotional line: free for the first two weeks after launch.

Counting from 5 November, that window closes around 19 November. Today it has roughly five days left. When it closes, the page says original prices resume and asks you to keep an eye on your balance, which is the entire pricing disclosure for a 560B flagship.

What you can do is bound the number from the row above. Ling-3.1-flash lists at ¥2.00 per million input tokens and ¥6.00 per million output, discounted 75% to ¥0.50 and ¥1.50, with cache reads at ¥0.40 falling to ¥0.10. Converting the list figure puts a text-only Ling flagship at roughly $0.28 input and $0.85 output per million; the discounted rate lands near $0.07 and $0.21. If Ling 3.2 Flash prices like its sibling, that is the neighbourhood. It is inference from an adjacent row, not a published rate, and a multimodal model carries image and video tokens whose pricing Ant has not described at all — a thirty-second clip at 2fps is 32 frames, and no page tells you what those 32 frames cost.

That gap has a specific consequence for anyone budgeting: you cannot answer "what does this task cost" for Ling 3.2 Flash the way you can for a model with a rate card. You can answer it after 19 November, or you can answer it now by running the evaluation while the meter is off — which is the sensible reading of a fourteen-day free window on a model with no published price.

Three benchmark numbers, all of them Ant's

Ant's announcement compares Ling 3.2 Flash against Opus 5 running at maximum reasoning effort on three evaluations. All three are vendor-reported, produced by Ant's own release pipeline, and none has been reproduced outside it. With that label attached:

• OmniDocBench1.5 — 92.21 for Ling 3.2 Flash against 90.56 for Opus 5 (max), a 1.65-point gap on document understanding.

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

• Vision2Web — 85.86 against 80.20, a vision-to-web-implementation evaluation.

Read the set and it says something narrower than "beats Opus 5." All three benchmarks are visual. Ant chose the three evaluations where a native-multimodal model at 25B active parameters could plausibly beat a frontier closed model, and published no text-only comparison — no reasoning suite, no coding suite, no agentic terminal task, which is precisely where the same lab's Ling-3.1-flash positioning lives. Two of the three, ZEROBench and Vision2Web, are not instruments with wide third-party participation, so a reader cannot sanity-check them against other models. And "Opus 5 (max)" is one configuration of one competitor, with no harness, prompt set or attempt count stated for either side.

There is also no independent score for this model to set beside them. Artificial Analysis, the usual neutral referee, does not have a page for Ling 3.2 Flash as of 14 November — the URL returns a 404 — and the closest thing the family has is its text-only sibling. Ling 3.1 Flash 1019 scores 29.6 on the current v5.0.1 Intelligence Index at $0.30 input and $0.90 output per million. That is a real independent number, and it belongs to the previous model, not this one.

The input limits are the real product spec

Ant documents four capability areas for Ling 3.2 Flash, and they are narrower than the usual launch list: complex visual understanding (documents with equations and diagrams, not captioning), multimodal long-horizon agents that keep a visual context alive across tool calls, 3D understanding and interaction, and medical multimodal work. The demos map onto those — reading headless renders of material spheres and comparing roughness to refine its own scripts, converting engineering drawings into 3D web models, coordinating specialist agents over a shared visual context to produce a report, a deck and a highlights video from the same Q2 data.

Then you read the limits, and the demos start to look like they were built inside the constraints rather than around them. A visual feedback loop that inspects its own renders is cheap in frames: one image per turn, up to forty images per request, each under 16384 by 784 pixels. A long-horizon agent that holds a meeting recording in context is not, because the recording is capped at thirty seconds and thirty-two frames. The two limits together describe a model built for many small looks at the world, not one long one.

Rate limits compound it. Ant's own rate-limit page sets a global ceiling of two requests per second, with 429s above it, and a 90-second timeout on synchronous calls with streaming recommended. For a chat product that is fine. For the multi-step agent loops the model is explicitly positioned for — dozens of calls per task, each carrying a frame — two QPS is a planning number, and it will shape architecture more than any benchmark gap will.

Where it fits, and what it costs you to try

The honest position on availability: Ling 3.2 Flash is not on the OrcaRouter catalogue today. The live model list returns 215 identifiers and none of them is a Ling model — the "ling" matches are Kling video models, which is a different Chinese lab entirely. So nothing here is one call away, and no failover path exists on our side for it. What we route is the comparison class it will be judged against: MiniMax M3.1 Flash, a 428B/23B native multimodal MoE with the same 1M window, video input and open weights, at $0.15 per million input and $0.60 per million output with a 90% cache discount — provider list price passed through at zero markup, with automatic failover across providers behind one key.

That is the useful shape of the decision while Ling 3.2 Flash has no rate card. Run the free evaluation on Ant's own API, because five days of free inference on a 560B multimodal model is worth the setup cost of a base-URL change. Do not put it behind a production path you cannot price, cannot fail over and cannot cap at more than two requests a second. If the workload is document or frame understanding and you need a number today, the models you can actually budget against are the ones on a rate card — and there are 200+ of them on one key, which at least makes the comparison a single integration rather than one per vendor.

What to watch before the window closes

Three things decide whether Ling 3.2 Flash becomes a model you use or a model you evaluated. The first is the rate card: whether Ant publishes per-million pricing — including how it bills image and video tokens — before 19 November, or whether the free window simply lapses into an unlisted rate. The second is weights. Ant is a lab whose 2026 reputation rests on MIT-licensed 560B checkpoints; Ling-3.1-flash-1019 shipped at 573.8 GB under MIT on 27 October, and there is no 3.2 repository on the inclusionAI Hugging Face organisation as of this week. A closed flagship from that lab is a posture change, not a formality. The third is the first independent benchmark, because until one exists every multimodal claim in this article is Ant's word against Ant's task selection.

None of that makes the model uninteresting. A 560B native multimodal MoE at 25B active, with a 1M window and both API dialects, is a serious piece of engineering, and the three benchmark gaps over Opus 5 — vendor-reported, unpublished harness — are large enough to be worth testing yourself rather than dismissing. The correct posture nine days in is the one the free window is built for: measure it now, price it when the rate card appears, and do not mistake a promotional fortnight for a product commitment.