Ant Group has launched Ling-3.2-flash, a native multimodal model with 560 billion total parameters that accepts text, images and video in a single system, the company said on Thursday. The model activates 25 billion parameters per token, supports a one-million-token context window, and is built for long-running agent work, visual understanding and medical imaging. It is available now through Ant’s chat interface and API, free for its first two weeks.
## What the model does
Ling-3.2-flash is the first model in Ant’s Ling line to handle video natively alongside text and images. Ant says it focuses on four capabilities: complex visual understanding, multimodal agents that sustain goals across multi-step tasks, three-dimensional scene understanding, and medical work such as reading scans and reports. The API works with both OpenAI and Anthropic protocols.
The size is the point. With 560 billion total parameters and only 25 billion active per token, the model follows the mixture-of-experts pattern: a large store of knowledge that spends a fraction of it on each step. That keeps inference costs down while holding a million-token window, enough to keep long documents, video and agent histories in view at once.
Ant’s own tests put the model ahead of Claude Opus 5 at its maximum setting on three benchmarks: OmniDocBench 1.5 at 92.21 against 90.56, ZEROBench at 43.71 against 39.52, and Vision2Web at 85.86 against 80.20. The three tests cover document understanding, visual reasoning and turning designs into web pages.
It also entered the DesignArena image-to-frontend ranking in the top ten with an Elo score of 1238, tied with models from Meta, DeepSeek and Anthropic. All of these are vendor-run numbers, and no independent lab has yet published a reproduction of them.
## Price and access
The API costs $0.30 per million input tokens and $0.90 per million output tokens, and it is free for the first two weeks after launch. It runs through Ant’s Bailian platform, and developers can try it in the chat interface before wiring it into their own systems.
## Our Read
Ant is joining a crowded week for Chinese models, and its entry is deliberately different. [Baidu opened the weights of its 280-billion-parameter ERNIE model this week](https://geroundtable.com/baidu-ernie-5-5-flash-open-weights/), and [Moonshot published Kimi K3.1 Pro’s weights](https://geroundtable.com/kimi-k3-1-pro-open-weights/), but Ant is keeping Ling-3.2-flash hosted behind an API.
The bet is that a cheap, capable hosted model wins developers who do not want to run their own infrastructure. The risk is that the benchmark claims are Ant’s own, and the model that matters most for agents — the one that sustains goals across long tasks — is the hardest kind to judge from a score table.
The comparison to watch is not with other Chinese models but with the frontier: [Thinking Machines’ Inkling 2 Small scored 31 on the Artificial Analysis index this week](https://geroundtable.com/thinking-machines-inkling-2-small/) with a similar agent-first pitch. If independent tests confirm even part of Ling’s vendor claims, Ant has a serious agent model at a bargain price.
## Why It Matters
A 560-billion-parameter model that takes video, holds a million tokens and costs $0.30 per million input tokens changes what a small team can build. Long-horizon agents need exactly that combination: enough context to hold a task, enough vision to read the world, and a price that survives a long rollout.
Ant’s free two-week window is an invitation to test whether the model delivers on that promise. The answer will come from developers, not from Ant’s benchmark table, and the first published independent evaluations will say more than any launch claim.
## Background
Ant’s Ling line has moved fast this autumn: Ling-3.0-flash-VL added multimodal support in September, and Ling-3.1-flash followed with long-horizon agent work. Ling-3.2-flash is the first in the line to take video input natively, a capability that previously belonged to Ant’s Ming series.
Source: [Ant Ling’s changelog](https://developer.ant-ling.com/en/docs/getting-started/changelog/) and [TechNode’s report](https://technode.com/2026/11/06/ant-launches-ling-3-2-flash-multimodal-model-with-million-token-context/)
0 Comments