Video, at 32 frames

Start with the number the launch does not lead with. Ant Ling's multimodal guide, the developer document that covers Ling-3.2-flash, caps video input at one clip per request, no longer than 30 seconds, sent as Base64 with an encoded size limit of 32 MiB. The model does not watch the clip. It samples it at a fixed rate of 2 frames per second and looks at no more than 32 frames.

That may be a reasonable engineering trade-off for some tasks. A static product demonstration, a slide recording or a surveillance sequence may retain much of its relevant information in 32 sampled frames. But whether the model actually captures that information reliably is a separate question, and the launch materials provide no benchmark results to establish it. For anything where the meaning lives between frames — a fast edit, a sleight of hand, a defect that flashes past in half a second — the sampling is the result. At two frames per second, the model can easily miss an event lasting a quarter of a second, depending on when it occurs relative to the sampling points. A 30-second ceiling also rules out longer recordings, including many of the extended sessions that agent workflows might need to inspect.

The 1M-token context window invites a bigger mental picture than the guide supports. A million tokens of text and a 32-frame look at one short clip are both real specifications of the same model. They are not the same kind of capacity, and only one of them belongs in a headline.

Images are roomier, within their own limits: up to 40 images per request, JPEG or PNG, Base64 only, with a 32 MiB ceiling on the request body.

Free for two weeks, priced at nothing published

The launch offer is clear: two weeks free from November 5. What replaces it is not. Separately, IA Decider's November 5 review reported that the model appeared at zero cost in the limited-time benefits section but had no entry in Ant Ling's regular model pricing table. The same review found no listing on OpenRouter, ZenMux or Novita, and no Ling-3.2-flash weights in the inclusionAI Hugging Face organization. Those are dated availability checks, not a guarantee that the model will remain absent from those services.

There is a precedent within the Ling catalog. IA Decider's review records that Ling-3.1-flash also launched with a free period before public pricing appeared, and that its weights subsequently became available under an MIT license. That history suggests Ling-3.2-flash could follow a similar path, but it does not establish when its weights will appear or what its eventual price will be. Until Ant Ling publishes a row for it, any cost estimate for production use is a guess, and the honest planning figure for week three is “unknown,” not “free.”

What a useful test looks like

The free window is enough time to learn the one thing the launch materials do not say: how well the model reads. Ant Ling's November 5 changelog entry publishes no benchmark scores. The launch materials therefore provide no quantitative basis for comparing Ling-3.2-flash with competing models. Independent testing will have to establish how well it performs. A test that will still mean something after the price arrives is simple. Use short clips where the answer is known, include at least one where the decisive moment is brief, and record how often the model reports what the frames it sampled could not have shown. That last failure mode, confident answers beyond the sample, is the one that decides whether a video-reading agent can be left unsupervised.

Sources

  1. [1] Ant Ling, developer changelog v2.0.6 (November 5, 2026; read first-hand at Gate 1: launch of the native multimodal Ling-3.2-flash — 560B total parameters, 25B activated parameters, 1M context window, native support for image, text and video input, four focus areas: complex visual understanding, multimodal long-horizon Agent, 3D understanding & interaction, and medical multimodal; available via the chat interface and APIs compatible with both OpenAI and Anthropic protocols; free for the first 2 weeks after launch; the entry publishes no benchmark scores)Read source
  2. [2] Ant Ling, Multimodal Understanding guide (read first-hand at Gate 1: the guide names Ling-3.2-flash as the latest-generation native multimodal model it covers, and states the video limits — formats MP4, MOV and WMV, one video per request, duration ≤30s, Base64 only with each encoded video ≤32 MiB, frame sampling fixed at 2 fps and capped at 32 frames per video — and the image limits: JPEG and PNG, up to 40 images per request, Base64 only, maximum request body 32 MiB)Read source
  3. [3] Ant Ling, official pricing page (rechecked first-hand at Gate 1 on November 6, 2026: Ling-3.2-flash appears in the Limited-Time Benefits as free for the first 2 weeks after launch, and has no row in the regular Model Pricing table or in the page's OpenRouter and ZenMux tables)Read source
  4. [4] TechNode, "Ant launches Ling-3.2-flash multimodal model with million-token context" (November 6, 2026; read first-hand at Gate 1: released November 5, a native multimodal model accepting text, images and video, 560 billion total parameters, 25 billion active, one-million-token context window, chat interface and OpenAI- and Anthropic-compatible APIs, free access for the first two weeks after launch)Read source
  5. [5] IA Decider, "Ant Ling Review 2026: Every Ling, Ring and Ming Model" (Ling-3.2-flash section checked November 5, 2026: the model billed at zero in the Limited-Time Benefits box, with no row in the Model Pricing table and no list rates published, no route on OpenRouter, ZenMux or Novita, and no 3.2 weights in the inclusionAI Hugging Face organization; Ling-3.1-flash background: launched with a free period before public pricing appeared, its weights subsequently available under an MIT license; Ling-3.0-flash-VL: 124B, launched September 4, 2026)Read source