Ant Ling's Ling-3.0-flash-VL adds 1M-token context and a separate 32-frame video cap

Multiply the documented video limits out and the frame ceiling arrives first. Thirty seconds sampled at two frames per second is 60 frames, and the documentation keeps 32 of them, which puts full-rate coverage at 16 seconds ^[9][1]. So either you send a 16-second clip and keep every sampled frame, or you send the full 30 and accept that something upstream throws away half of them. The frame budget, not the million-token context window described in the architecture ^[8], is the binding limit on video work.

The image path has its own envelope. Requests take JPEG and PNG, up to 40 images, inside a 32 MB maximum request body ^[10]. Fill all 40 slots and each image gets 0.8 MB on average ^[2]. An agent that carries a scrollback of screenshots in one request is therefore trading image count against image sharpness, and small interface text is the first thing to go when you downscale. Ant Ling's own framing points the same way: the company's claims about searching a timeline and extracting keyframes appear to rely on an agent splitting the job into tool calls rather than putting a long recording into one request ^[15].

The 38 on the Intelligence Index belongs to the base model, and Artificial Analysis measured it, along with the 5.1 billion active parameters per token ^[3][4]. The 42 belongs to Ant Ling, which attributes the gain to training visual and language capabilities together ^[5]. As of September 4, Artificial Analysis had published no page for the VL model ^[6]. For those four points to transfer to your own shortlist, the party that produced the 38 would have to score the VL model on the same index at the same version, and the mechanism would have to be that image training moved a text score. That is a testable claim, and on this evidence nobody outside Ant has tested it.

The sparsity is the part of the design that looks well aimed. About 4 percent of the parameters do work on any given token ^[3], which is a serving-cost story for whoever holds the weights and a latency story for everyone else. Whether you can be the former is not settled by this release: it documents the vision transformer, the two-layer MLP projector and VideoRoPE positional encoding ^[7], the hybrid backbone ^[8] and the request limits ^[9][10], and it states no licence or weights download for Ling-3.0-flash-VL ^[4]. Ant Group's open-source record under Richard Bian ^[16] provides useful context, though it stops short of a licence file for this release.

The demonstration that will get quoted is the one where the model takes a design reference, writes the site, opens it in a browser, compares the render against the reference and revises the code ^[11]. Ant Ling also says a single reference image can steer layout, color, typography and interface components across a new site ^[13]. These are company demonstrations, not independent evaluations ^[14]. The loop they describe spends one screenshot per iteration, which places the 40-image, 32 MB envelope directly across the feature the release is built around, and that is the first number I would test rather than the index score.