On November 5, Ant Ling launched Ling-3.2-flash, a natively multimodal model with 560B total parameters and 25B activated per step, a 1M-token context window, and native image, text, and video input. The point of this upgrade is not the parameter count — it is vision built into the reasoning base, so the model can take on long-horizon multimodal tasks directly.
560B parameters, only 25B active: how does the compute math work?
Like Ling-3.1-flash, released in September, Ling-3.2-flash is a mixture-of-experts model with 560B total parameters and about 25B active: at inference time it only calls the expert networks relevant to the current token, using far less actual compute to wield a much larger model. The difference is that 3.1 was a hybrid model focused on reasoning, while 3.2 builds vision into the architecture itself instead of bolting a vision module onto a language model. The 1M context lets it swallow a long document, a long video, or a multi-turn conversation in one pass and then reason across modalities. For developers, that means lower inference cost for the same task — but only if the task genuinely needs image, text, and video together. For pure text work, it is a waste.
It beats Opus 5 on three benchmarks — how much should that count for?
In Ant's published vendor benchmarks, Ling-3.2-flash beats Claude Opus 5 (max) on three tests: 92.21 versus 90.56 on the document-understanding benchmark OmniDocBench 1.5, 43.71 versus 39.52 on a ZEROBench subset, and 85.86 versus 80.20 on Vision2Web, which turns images into websites. On the third-party DesignArena image-to-frontend leaderboard it ranks in the top ten with an Elo of 1238. But all of these numbers are vendor-reported and not yet independently verified: the benchmarks are tracks that favor the model, and the full test setup for Opus 5 has not been disclosed. What will actually settle it is feedback from developers on real tasks after launch, not three score tables on release day.
Can you use it now, and will the weights be open-sourced?
The model is already available to try in the Ant Ling chat interface, and its API is compatible with both OpenAI and Anthropic protocols, free for the first two weeks. IT Home reports API pricing of $0.30 per million input tokens and $0.90 per million output tokens. Ant has not said whether Ling-3.2-flash weights will be open-sourced: the previous generation, Ling-3.1-flash, released its weights under the MIT license after the free period ended, and whether 3.2 follows the same path depends on what Ant says next. Teams in a hurry can use the free window to validate it; anyone planning a local deployment should not wait — treat the open-source 3.1 as the fallback.
