Ant Ling opens its vision model under MIT: 124B parameters, 5.5B active per token
Ling-3.0-flash-VL is out as open source. The mixture-of-experts architecture holds 124 billion parameters but activates only 5.5 billion per token, with a context window up to 1 million tokens. It also claims to act on what it sees on screen.
Ant Ling, the AI team at Ant Group, has open-sourced its vision-language model Ling-3.0-flash-VL under the MIT licence. It is Ling-3.0-flash, released in July, with sight added.
The numbers
The architecture is mixture-of-experts: 124 billion parameters in total, of which only 5.5 billion activate per token. The context window reaches 1 million tokens. The backbone is a 42-layer hybrid that alternates KDA and Gated MLA layers at a 5:1 ratio. On the vision side a ViT encoder extracts features from images and video, a two-layer MLP projector aligns them with text, and VideoRoPE encodes both spatial position and temporal order. Published formats are BF16 and F32; the announcement said FP8 shipped too, with FP4 and INT4 to follow.
Sparse activation is the point here. A 124B model that runs 5.5B per token costs far less to serve than a dense model of the same size. For an open-source release, that is what closes the gap between downloading a model and actually running it.
Three jobs
The model card groups the capability under three headings. Understand: counting objects, parsing layouts, reading charts and documents. Reason: calculating from visual evidence, multi-step inference, verification. Act: understanding interfaces and completing tasks, meaning it looks at web and software screens and sequences actions.
The third is the interesting one. Recognising what is on a screen and deciding what to do on that screen are separate jobs; the second is agent work, and it is still thin on the open-source side.
The benchmark
The only comparative figure on the card is 42 on the Artificial Analysis Intelligence Index v4.1.1, four points above the 38 scored by the sightless Ling-3.0-flash. There is no line-by-line competitor table, so where this model stands cannot be stated until independent measurements arrive.
Ant Ling's models were also in the news with the health-focused Ling 3.0 Flash Sante, which we covered on 6 September; the team keeps branching specialised models off the same trunk.
