Ant Group has released Ling-3.2-flash, a new multimodal artificial intelligence model that accepts text, images and video as input, the company's Ant Ling team confirmed on 5 November 2026.

The model is built on a mixture-of-experts design with 560 billion total parameters, of which around 25 billion are activated for each inference step. It supports a context window of one million tokens, allowing it to work with very long documents, extended conversations and large sets of visual material in a single session.

Native multimodal design

Unlike models that bolt a vision system onto a text model, Ling-3.2-flash is described by Ant as natively multimodal, meaning image, text and video are handled within the same architecture. The company says the release focuses on four areas: complex visual understanding, multistep agent tasks that run over a long horizon, three-dimensional understanding and interaction, and medical multimodal work.

The model is available through an online chat interface and through an application programming interface that works with both OpenAI and Anthropic protocols, making it easier for developers to switch existing tools over to it. Ant is offering free access for the first two weeks after launch. Reported API pricing after that period is $0.30 per million input tokens and $0.90 per million output tokens.

Vendor benchmark claims

Ant says Ling-3.2-flash scored above Anthropic's Claude Opus 5 (max) on three of its own release evaluations: OmniDocBench1.5, where it scored 92.21 against 90.56; a ZEROBench subset, at 43.71 against 39.52; and Vision2Web, at 85.86 against 80.20. Those are vendor-run tests, not independent benchmarks, so they should be treated as the company's own claims until third parties reproduce them.

The model also entered the top ten of the DesignArena image-to-frontend ranking with an Elo score of 1238, placing it alongside models from Meta, DeepSeek and Anthropic in that arena.

China's crowded AI race

The launch adds to a fast-moving week for Chinese AI labs. Ant's Ling line has moved quickly through 2026, from the Ling-2.6 flagship models in the spring to the Ling-3.0 family in the summer and now the 3.2 generation. The emphasis this time is less on raw size than on practical use: agents that can see, plan across many steps and act on what they observe.

Ant has also indicated that the model weights will be opened up in the near future, although the launch itself is through the chat interface and API rather than an open-weights release. For developers, the two-week free window and dual-protocol compatibility lower the cost of trying the model on real workloads, from document-heavy visual tasks to agent-style automation.

Whether Ling-3.2-flash earns a lasting place in production systems will depend on independent testing and on how it performs outside Ant's own evaluation suite. For now, it is another sign that China's AI labs are competing directly with the leading Western models on multimodal capability, not only on price.