Weights are not out, and the 3D gap is stated

The Ant Ling post is timed 19:04 in Beijing on 6 November 2026, which is 2026-11-06T11:04:00+00:00. The post says a native multimodal model already running in Lingguang 2.0 is being introduced that day under the name Ling-3.2-flash. The stated size is 560 billion total parameters and 25 billion active parameters per step, with a native context of 1 million tokens.

The post puts the emphasis on checking results. It says the model can take an intermediate render or an interface as a new input, compare differences in geometry, light, and layout, and then adjust code or steps. It also says that when execution fails, a syntax error appears, or a tool call fails, the model uses the error context and the stack log to retry or change the branch. That is the company's description of a capability. This column does not turn those sentences into steps to follow.

The only specific score written in the prose is 1238 on DesignArena's Image-to-Frontend code-generation evaluation. The post calls that score a reference for turning a visual design into a front-end implementation, and it does not state a rank. The post also names categories it says were observed, including visual reasoning, document parsing, medical multimodal tasks, and agent interaction, without printing scores for those categories in the prose. Medical appears only as a category name. The post does not describe a treatment or an experiment procedure, and this column does not either.

The speed figures are the post's own measurements. The architecture is described as a 7-to-1 mixed attention stack: seven KDA layers and one Gated MLA layer. Under the speculative method the post calls DSpark, with BlockSize set to 5, the post says the average accepted length exceeds 3.5 on both multimodal and text tasks. The stated conditions are a Hopper GPU and a context of about 30,000 tokens, with a mean decode speed of 280 tokens per second for a single user and about 3.54 milliseconds per output token. The post attributes those speeds to production data.

The post states the gap itself: long-running cross-application state, large-scale dynamic reasoning in 3D space, and the design taste and detail fidelity of 3D assets still trail what it calls leading closed models. The try-it links are Ling studio and one model-service page on Ant Group's digital-technology platform. The weights are described as opening soon. The post gives neither a date nor a repository address.