|
Ant Group's Ling team has launched its new native multimodal model. (Image: TechNode)
Is China's payments giant now an AI lab too? Ant Group's Ling team says yes, and on November 5 it shipped the proof: Ling-3.2-flash, a native multimodal model carrying 560 billion total parameters while activating only 25 billion per token, with a one-million-token context window and built-in handling of text, images and video, as reported by TechNode and TLDR.
Ant positions the release around four strengths: complex visual understanding, long-horizon multistep agent work, 3D understanding and interaction, and medical multimodal analysis. Its published evaluations claim wins over Anthropic's Claude Opus 5 on OmniDocBench1.5 (92.21 to 90.56), ZEROBench Sub (43.71 to 39.52) and Vision2Web (85.86 to 80.20). If independent tests agree, a Chinese open-leaning model now leads the document-understruction pack, and developers get a serious alternative for vision-heavy workloads.
What makes the API setup notable?
The model speaks both OpenAI's and Anthropic's API dialects, so a team can point existing code at it with minimal changes, and Ant is making it free for the first two weeks, per TechNode. For builders in Nigeria and other cost-conscious markets, that means trialling a frontier-class multimodal model costs one afternoon, not a procurement process.
Why does the 560B/25B split matter?
It is the mixture-of-experts economics the industry now relies on: huge stored knowledge, tiny per-token compute bill. The Ling family also includes the Ring reasoning line and the Ming any-to-any line, so Ant is building a complete stack rather than betting on a single release.
The wider signal is competition. Ant, best known for Alipay, now spends at frontier scale alongside Alibaba, ByteDance and DeepSeek, and every comparable release keeps pressure on pricing across the AI industry. That pressure is what eventually reaches the small studios and solo developers whose budgets decide which models actually get built on.
For Nigerian developers and small teams, the practical moves this week: open the free trial while it lasts, test document and vision tasks, the model's claimed strong suit, and keep your AI plan multi-provider so a price or policy change at one vendor never strands your product.
FAQ
What did Ant Group launch? Ling-3.2-flash, a native multimodal model with 560B total parameters, 25B active per token, a 1M-token context and text, image and video input.
Is it free? The API is free for the first two weeks after launch, per TechNode; paid tiers follow.
Can I drop it into an existing app? Yes: OpenAI- and Anthropic-compatible APIs mean most integrations need only an endpoint change.
Sources: TechNode, TLDR, Ant Ling official site, AI Weekly.
|
0 Comments