Ant Group's AI lab has released Ling-3.2-flash, a native multimodal model that handles text, images and video in a single architecture — and, on the company's own benchmarks, beats Anthropic's Claude Opus 5 on three vision tests. The model, announced on November 5, is the latest sign that Chinese labs are closing in on, and in places passing, Western frontier models on multimodal work.
Ling-3.2-flash is a mixture-of-experts model with 560 billion total parameters and 25 billion activated per inference step, with a one-million-token context window. It is available now through Ant's Ling Studio chat interface and its model-service platform, with APIs compatible with both OpenAI and Anthropic protocols, and free access for the first two weeks after launch.
Vision built in, not bolted on
Unlike models that add a vision encoder to a text model, Ling-3.2-flash natively integrates visual and language processing at the architectural level, using a 7:1 hybrid attention design. It supports image, text and video inputs, and Ant says it focuses on four capabilities: complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work such as interpreting images and generating reports.
Its headline feature is visual feedback self-inspection: the model can use images and interactive interfaces as observational inputs, compare differences in structure, lighting and layout, and automatically correct its own output. When it hits errors or tool-call failures, it can locate the problem, retry and compensate, keeping long tasks moving. Ant has deployed the model in Lingguang 2.0, its agent product aimed at reliable execution in real production environments.
The Opus 5 claim — with a caveat
In Ant's own published evaluations, Ling-3.2-flash surpasses Claude Opus 5 (max) on three benchmarks: OmniDocBench 1.5, where it scores 92.21 against 90.56; ZEROBench (Sub), 43.71 against 39.52; and Vision2Web, 85.86 against 80.20. It also entered the top ten of DesignArena's image-to-frontend arena with an Elo score of 1238, tied with models from Meta, DeepSeek and Anthropic, and DesignArena rates its pricing — $0.30 per million input tokens and $0.90 per million output tokens — as the lowest in its performance tier.
The caveat matters: these are vendor-run evaluations, not independent third-party tests, and Ant has not yet put the model through a widely used independent index. The company itself acknowledges shortcomings in long-range planning and 3D asset detail. On speed, Ant reports average single-user decoding of 280 tokens per second on Hopper-architecture GPUs at around 30,000 tokens of context, with latency of about 3.54 milliseconds per token.
China's open-model wave keeps rolling
Ling-3.2-flash lands in a crowded week for Chinese AI. Kimi's K3, DeepSeek's Harness agent platform and Tencent's Hy4 have all made headlines in recent days, and Ant's own Ling-3.0-flash-VL shipped open weights earlier this year. This time, though, there are no weights at launch — Ling-3.2-flash is proprietary for now, though Ant says the weights will be opened up in the near future.
For Ant Group, the company behind Alipay, the model is also a commercial play: cheap, fast multimodal inference aimed at developers building agents that see and act, from turning design drafts into working web pages to operating professional software like Blender. Whether its Opus-beating scores hold up under independent testing will decide how seriously the wider industry takes the claim — but on price and speed alone, Ling-3.2-flash is already hard to ignore.
