Ant Ling launches Ling-3.2-flash native multimodal model
Ant Ling released Ling-3.2-flash, a native multimodal MoE model with 560B total parameters, 25B activated parameters and a maximum 1M-token context window. It accepts text, image and video inputs. Ant Ling’s official evaluation reported that the model outperformed Claude Opus 5 (max) on OmniDocBench 1.5, ZEROBench (Sub) and Vision2Web, but a supplied commentary said these were vendor-selected results without independent third-party evaluation at launch. The API is available through Ant Bailian, supports OpenAI- and Anthropic-compatible protocols, and is priced at $0.30 per million input tokens and $0.90 per million output tokens. The first two weeks of access are free.
The release adds an API-based option for developers working with image, text and video inputs and long contexts, while the reported benchmark advantage over Claude Opus 5 remains unverified by independent evaluations.
TL;DR:11 月 5 日,蚂蚁集团 AI 团队蚂蚁百灵发布原生多模态大模型 Ling-3.2-flash——560B 总参数、单步只激活约 25B 的 MoE 架构,1M Token 上下文,图像/文本/视频直接进模型。API 同时兼容 OpenAI 和 Anthropic 协议,输入定价 $0.30/百万 token(约 2 元人民币),前两周免费。官方自评三项基准超过 Claude Opus 5(max);不过这次没有开源权重,独立第三方评测也还没跟上。
给 agent 应用开发者提个醒:你的多模态 API 选型清单,这周可以多写一行候选了。
11 月 5 日,蚂蚁百灵在开发者平台更新日志里放出了 Ling 系列的新一代旗舰 Ling-3.2-flash。名字里的 "flash" 不是营销词——它指这个模型的「参数经济学」:总参数 560B(约 5600 亿),但每个 token 前向计算只激活约 25B(约 250 亿),激活比例 4.46%。打个比方:一家医院有海量专科医生(总参数决定知识容量),但接诊时只调度与病情相关的科室(激活参数决定单次成本)。
#Ant Ling#Ling-3.2-flash#Ant Bailian#Claude Opus 5#Anthropic#OpenAI#DesignArena