Ant's 560B image-and-video model is priced at US$0.30 per million input tokens, the lowest among comparable models according to DesignArena, as reported by IT Home.
Ant Group released Ling-3.2-flash on November 5, a model that accepts text, images and video and keeps the 560 billion total parameters, 25 billion of them active per step, and the 1-million-token context window of its text-only predecessor. Ant’s changelog lists the release under version v2.0.6, and TechNode and IT Home reported it on November 6. A token is the small unit of text a model reads and writes; a context window is how much of it the model can consider at once.
Ant says it built the model for complex visual understanding, multistep agent tasks, 3D understanding and interaction, and medical multimodal work. It is available through an online chat page and through programming interfaces compatible with both OpenAI’s and Anthropic’s protocols, and Ant says access is free for the first two weeks after launch. IT Home lists API prices of US$0.30 per million input tokens and US$0.90 per million output tokens, which it converts to roughly RMB2 and RMB6.
The vision benchmark comparisons are Ant’s own, and none of the outlets reports an outside evaluation of them. Ant has also not said, in the sources reviewed, whether it will publish the weights, the downloadable files that let others run a model themselves. The release extends Ling-3.1-flash into vision, and IT Home says the DesignArena benchmark site lists the price as the lowest among models of comparable performance.
Ling-3.2-flash keeps the size of Ling-3.1 and adds sight
The two models share the same headline specifications. Ling-3.1-flash, which EastFrontier covered when it appeared on September 30, has about 560 billion total parameters and 25 billion active per token, according to its Hugging Face model card. The card lists the model as text generation under the MIT license, with a 1-million-token context. Ling-3.2-flash adds native image and video input to the same scale, and IT Home says a 560B total and 25B active split typically corresponds to a mixture-of-experts design, in which only a fraction of the model runs for each step, and reads the combination as a trade-off between capability and inference cost. The Ling-3.1-flash card describes that model as a mixture-of-experts system that routes each token to 8 of 512 experts.
Ant’s changelog entry for Ling-3.2-flash lists the model as natively accepting image, text and video input with a 1-million-token window, and it names no limit on that window during the two-week free period. The changelog gives no price and no benchmark results, so the figures in this story that go beyond the specifications come from IT Home and DoNews.
Ant’s benchmark wins come from three vision tests it ran itself
IT Home and DoNews both carry the comparison with Claude Opus 5 (max), as they label the Anthropic model. Ant’s own tests put Ling-3.2-flash ahead of Claude Opus 5 (max) on three vision benchmarks, including 92.21 against 90.56 on OmniDocBench1.5. That test measures how well a model reads documents. The other two are ZEROBench (Sub), where the scores are 43.71 against 39.52, and Vision2Web, where they are 85.86 against 80.20.
IT Home adds that the model scored an Elo rating of 1238 on DesignArena’s image-to-frontend arena, which places it in the top ten alongside models from Meta, DeepSeek and Anthropic. DoNews does not say how the Opus 5 comparison was set up, and neither outlet reports results on text-only benchmarks, so the comparison says little about whether Ling-3.2-flash also improves on Ling-3.1-flash’s text reasoning.
Whether Ant publishes the weights will decide who builds on the model
Ant has published Ling-3.1-flash’s weights on Hugging Face. Its organization page there, inclusionAI, describes itself as the home of Ant Group’s artificial general intelligence initiative and releases projects under MIT or Apache licenses. The page lists 234 models but displayed only 10 of them, including Ling-3.1-flash and Ling-3.1-flash-1019, and no Ling-3.2-flash entry appeared among those shown, so the new model’s weights may exist in the part of the list the page did not display.
When the free period ends, developers will weigh the reported US$0.30 input price against the open Ling-3.1-flash weights they can already run themselves. The signals to watch are a Hugging Face model card for Ling-3.2-flash, independent results on the three vision benchmarks, and the first pricing Ant publishes beyond the free period.