One image as the brief, agents to implement it

Ant Ling’s Ling-3.2-flash model uses visual references to drive tools, create deliverables, and correct their rendering.

A reference image, then successive renderings to adjust the materials and surface details: this is one of the demonstrations chosen by Ant Ling to present Ling-3.2-flash. The new multimodal model is described as capable of using what it sees to carry out a task, manipulate software, and resume its work after inspecting the result. Image reading becomes a step in the creation process.

The published examples concern 3D creation and web interfaces, but also document production. What they have in common is project tracking: interpreting a reference, acting in a tool, then examining what has been produced. 3D renderings to guide corrections

In the scenario devoted to objects for gaming and simulation, a single 2D image serves as a reference. The agent then examines renderings of material spheres to compare their roughness and surface details. It gradually modifies the scripts based on these observations. The visual quality of the result here depends on a succession of trials and corrections, rather than a single proposal.

Controlling Blender illustrates another form of intervention. According to Ant Ling, the agent acts within the software interface and changes its settings; another example turns technical drawings into 3D models intended for the web. These achievements go through tools and code. They should not be confused with a 3D output directly generated by the model itself.

The same work from previews appears in brochure layout. Company documents provide the initial material, then the system visually checks the arrangement of the elements and their spacing to refine them. For web development, the demonstrations describe interfaces adapted to several screen sizes and featuring animations. Visual feedback therefore serves as much to organize a page as to work on the appearance of an object. Multiple agents around the same file

Collaboration between agents takes as an example second-quarter financial data. Specialized agents, coordinated around a shared context, produce an analysis report, a six-slide presentation for a board meeting, and a summary video. The deliverables are then revised at the user’s request.

The exercise focuses on adapting the same file into multiple formats. It combines content analysis with the creation of materials, then their revision after feedback. The launch thread does not specify either the duration of this assignment or its total cost.

Scientific documents constitute another source material. According to the team, the system reads equations and diagrams from research PDFs to translate the methods presented into executable code. This example mobilizes visual understanding within the document itself, whereas reading only the text would leave out part of the necessary information. A large context and API access

The model documentation announces 560 billion parameters in total, including 25 billion activated during token processing. This mixture-of-experts, or MoE, architecture only mobilizes part of the model at each step. The number of active parameters therefore does not represent the total size of the system.

The announced maximum context reaches one million tokens, with support for text, images, and video as input. Access is offered through a chat interface and via API, with protocols compatible with those of OpenAI and Anthropic. The launch thread also points to the model page on Computrix so it can be tried with one’s own images.

A launch offer provides free access during the first two weeks. The pricing page then announces a return to the usual prices, but does not yet give a specific price for this version in its table. Three scores against Opus 5 in Ant Ling’s evaluations

The results published at launch place Ling-3.2-flash ahead of Opus 5 in max configuration on three tests: 92.21 versus 90.56 on OmniDocBench 1.5, 43.71 versus 39.52 on ZEROBench (Sub), and 85.86 versus 80.20 on Vision2Web. These figures are those communicated by Ant Ling, and not an independent comparison.

The mention “Sub” specifically indicates the use of a subset for ZEROBench. The launch thread does not specify its composition or the full set of reproduction conditions. These scores also do not give the success rate of the projects presented: completing a scene in Blender or revising multiple deliverables requires chaining together operations that these three values do not describe.