November 6 news: The Ant Bailing official team today officially announced the launch of a new native multimodal large model — Ling-3.2-flash. The model has 560B total parameters, 25B activated parameters, and an ultra-long 1M context window. At the architecture level, it deeply integrates vision and language capabilities, bringing a new option for multimodal AI applications.
Four major capability upgrades, outperforming top industry players on multiple benchmarks
In its architecture design, Ling-3.2-flash natively integrates visual and language processing capabilities from the ground up and fully supports image, text, and video inputs. Official materials show that the model focuses on strengthening four core dimensions:
Complex visual understanding: deep reasoning and analysis for images and videos.
Multimodal long-horizon Agent: ensures the model stays goal-consistent across multi-step complex tasks and efficiently completes decision-making and execution by combining visual and textual information.
3D understanding and interaction: focuses on interpreting 3D scenes and spatial relationships.
Medical multimodal: deeply covers medical imaging understanding and report interpretation, providing effective assistance for professional medical analysis.
In evaluation results released by Ant, the model successfully surpassed Anthropic Claude Opus 5 (max) in three key benchmarks:
OmniDocBench 1.5: scored 92.21, compared with the competitor's 90.56.
ZEROBench (Sub): scored 43.71, compared with the competitor's 39.52.
Vision2Web: scored 85.86, compared with the competitor's 80.20.
In addition, in DesignArena's image-to-frontend arena, Ling-3.2-flash made it into the top 10 with an Elo score of 1238, tied with leading industry competitors such as Meta, DeepSeek, and Anthropic.
Architecture trade-offs and a highly competitive pricing strategy
In terms of technical implementation, the combination of 560B total parameters and about 25B activated parameters is the model's core trade-off between strong capability and inference cost. This mixture-of-experts (MoE) architecture ensures that the model invokes only part of its parameters during each forward computation, thereby keeping per-step compute cost within a reasonable range while retaining a huge knowledge reserve.
In terms of API pricing and experience, Ling-3.2-flash's API input price is $0.30 per million tokens (about RMB 2), and its output price is $0.90 per million tokens (about RMB 6). The DesignArena leaderboard notes that this is the lowest rate among models at the same performance level. Currently, the model's API has officially launched via the Ant Bailian platform. It is not only fully compatible with OpenAI and Anthropic protocol specifications, but also offers users a limited-time free trial for the first two weeks after launch.
