Ling-3.2-flash adds image and video input without changing its headline parameter count

Ant’s Ling-3.2-flash brings image and video input to the same large-model scale as Ling-3.1-flash: 560 billion total parameters, about 25 billion active per token and a one-million-token context window. The November 5 developer changelog presents it as a native multimodal model, available through the company’s API.

The change is clearer when compared with the family’s existing roles. Ling-3.1-flash already occupied the 560B/25B, million-token tier. The earlier Ling-3.0-flash-VL offered visual input at 124B total and 5.5B active parameters, with a 256,000-token context. The new release combines the larger tier with multimodal understanding; it is not simply the older visual model with a different name.

Three numbers answer three different questions

Total parameters describe the overall model’s scale. Active parameters describe the subset used for a token in a mixture-of-experts architecture. Context length describes the amount of input and conversation the system can accommodate, subject to its implementation and limits.

None is a direct quality score. A 25B active-parameter count does not mean the entire 560B model fits into the memory needed for an ordinary 25B dense model. Nor does a million-token window establish that every relevant detail will be recovered correctly from a very long document or video.

The figures come from Ant’s model documentation, not independent measurements. They establish the advertised configuration and a useful family comparison, while leaving task-level reliability to evaluation.

Visual input can change what an agent can check

An assistant working only from text can read a description of an interface. A system that accepts images can also inspect the rendered interface itself. Video adds a sequence of observations, potentially exposing changes that a single frame misses.

That is why the practical distinction is more interesting than the parameter increase alone. In a hypothetical interface-building workflow, a screenshot could reveal that a label is clipped even when the code compiles. A video could show a control changing state after a click. These are examples of what multimodal input makes possible, not demonstrations that Ling completes either task reliably.

Ant highlights complex visual understanding, long-running multimodal agents and 3D understanding. Its documents also discuss medical applications, but an advertised application area does not establish clinical validity. The release material should not be read as a replacement for domain-specific validation or professional judgment.

API compatibility lowers integration work, not every difference

The changelog lists compatibility with OpenAI and Anthropic API protocols. That can reduce the amount of request-format work needed to try the model in an existing application. It does not establish identical tool behavior, limits, output quality or data-handling conditions across providers.

The launch includes a two-week free-use promotion. Treat that as a dated introductory offer rather than a permanent pricing policy, and check the developer console for the applicable terms. The model-family documentation’s discussion of open releases should also not be generalized into a claim that this specific checkpoint has downloadable weights unless its release page supplies them.

For a developer, the useful opportunity is to test a broader input set without abandoning familiar integration patterns. The decision should turn on a representative task: the screenshot understood correctly, the relevant video moment found, the tool result checked. A million-token limit provides room for evidence; it does not decide what the model will make of it.