Ant Ling Launches Ling 3.2 Flash, a 560B Multimodal Model It Says Beats Claude Opus 5 on Release Evals
Introducing Ling-3.2-flash, Ant Ling's native multimodal model: 560B total / 25B active parameters. Stronger complex visual understanding, long-horizon multimodal agents, 3D understanding & interaction, and medical intelligence. Outscores Opus 5 (max) in release evaluations: - OmniDocBench1.5: 92.21 vs 90.56 - ZEROBench(Sub): 43.71 vs 39.52 - Vision2Web: 85.86 vs 80.20
Context
A cheap native multimodal MoE from Ant Group expands Chinese options for vision, 3D and agent work.
Ant Ling (@AntLingAGI) is Ant Group's AI lab brand. Ling 3.2 Flash is presented as a native multimodal mixture of experts model. The release evals compare against Anthropic Claude Opus 5 (max) on vendor chosen benchmarks and should be treated as company figures until independent runs appear.
Deep dive
Ant Ling (@AntLingAGI) posted on Nov 5 introducing Ling 3.2 Flash, calling it a native multimodal model with 560 billion total parameters and 25 billion active per step.
According to the company post, the model targets complex visual understanding, long horizon multimodal agents, 3D understanding and interaction, and medical intelligence. Ant Ling says its release evaluations beat Claude Opus 5 (max) on OmniDocBench1.5 (92.21 vs 90.56), ZEROBench Sub (43.71 vs 39.52), and Vision2Web (85.86 vs 80.20). Those scores are vendor reported.
Coverage around the launch also cites an API with about two weeks free and pricing near $0.30 input and $0.90 output per million tokens, plus OpenAI and Anthropic compatible endpoints. Weights are not described as open in the launch post.
The timing lands beside OpenAI's GPT 6 Cyber and Google's Gemini Omni 1.2 Flash the same day, so buyers will compare real multimodal agent demos, not only the chart. Some commentators noted Opus 5 is no longer Anthropic's newest flagship, which softens the headline comparison.
The lead post passed about 127,000 views.
What to expect next
Watch for independent benchmarks against Claude Opus 5.5 and GPT 6 class models, and for whether Ant Ling opens weights. Early API users will test document, video and 3D workflows that the launch demos advertise. Price and latency will decide if Ling 3.2 Flash becomes a default multimodal backend outside China. Any open weight follow up would change the competitive picture quickly.