Ant Ling releases Ling-3.0-flash-VL for visual agents and GUI control
Ling-3.0-flash-VL reads images and short videos, writes websites from screenshots, and operates interfaces through Ant Ling's hosted API.
By RuntimeWire Staff · Published
Primary source: Ant Ling
Why it matters
Ant is turning its efficient text model into an agent that can see and operate software, a direct bet that visual feedback will make coding and GUI automation more reliable.
Richard Bian (@RichSFO), Ant Ling's product and growth lead, has spent years translating Ant Group's open-source ambitions into products developers can use. On September 4th, his team released Ling-3.0-flash-VL, a version of its fast language model that accepts images and video and can act on what it sees. The launch moves Ant Ling beyond text generation and into visual agents that write interfaces, inspect their own work and operate software.
Ant Ling's Ling-3.0-flash-VL release announcement
The model is built on Ling-3.0-flash, the 124 billion-parameter mixture-of-experts model Ant introduced in July 2026. Artificial Analysis independently gives that base model a score of 38 on its Intelligence Index and measures 5.1 billion active parameters per token. Ant Ling says the visual version raises the score to 42, attributing the improvement to training visual and language capabilities together. Artificial Analysis had not published a corresponding page for the VL model as of September 4th, so the four-point gain remains Ant Ling's claim.
That distinction matters because Ant is positioning visual input as an intelligence upgrade, rather than a feature bolted onto an otherwise unchanged model. The demonstrations show a system meant to close the loop between perception and action: Ling-3.0-flash-VL can receive a design reference, generate a website, render it in a browser, compare the result with the reference and revise the code. Another demo has the model identify interface elements and complete tasks using clicks, typing, scrolling and application switching.
The model is built to inspect its own output
Visual feedback is the most consequential part of the release for developers. Coding models have become competent at producing a first draft from a text prompt, while verification still depends heavily on test suites or a person looking at the screen. Ant Ling is betting that a model able to inspect rendered output can handle part of that review cycle itself.
In Ant Ling's demonstration, the model receives a screenshot or design reference, writes the site, opens it in a browser and makes revisions after comparing the two versions. Ant Ling also says one reference image can steer layout, color, typography and interface components across a new site.
Ant Ling's coding-with-visual-feedback demonstration
Those examples are company demonstrations, not independent evaluations. They still reveal where Ant wants Ling to compete. The product is aimed at workflows where an agent must observe a changing interface and decide what to do next, rather than simply return code or text.
The technical design follows that objective. Ant Ling says a vision transformer extracts features from images and video, while a two-layer MLP projector aligns those features with the language model. VideoRoPE encodes spatial and temporal position. The company describes a 42-layer hybrid backbone that mixes KDA and Gated MLA layers at a 5:1 ratio and supports a 1 million-token context window.
The production limits are narrower than that context figure may suggest. Ant Ling's developer documentation caps a video at 30 seconds, samples it at two frames per second and processes no more than 32 frames. Each request can contain one video. Image requests support JPEG and PNG files, with as many as 40 images in one request and a 32 MB maximum request body.
A million-token context can accommodate large text and multimodal workloads, but it does not give developers unrestricted video analysis. For now, the documented interface is designed around short clips, selected frames and compressed visual evidence. Ant's broader claims about searching a timeline and extracting keyframes appear to rely on an agent breaking the job into tool calls rather than placing a long recording directly into one model request.
Bian's open-source operation becomes a model business
Bian arrived at Ant Ling after building Ant Group's open-source software organization. In a November 2025 interview with Interconnects, he said he had turned a one-line technical strategy into an eight-person, cross-functional operation over four years. Before Ant, he spent 11 years in the United States, including engineering roles at Microsoft and Square, and completed an MBA at the University of California, Berkeley's Haas School of Business.
That background helps explain Ant Ling's developer-first packaging. The API accepts text, image and video inputs through an OpenAI-compatible interface. Ant has also published integration instructions for coding and agent tools, making the model easier to test without requiring developers to adopt a proprietary request format.
Bian described inclusionAI in 2025 as an effort inspired partly by DeepSeek's open research and partly by Ant Group's need to build core AI capabilities for its financial, insurance and technology businesses. InclusionAI says Ant Group founded the lab and staffed it with researchers who previously worked at OpenAI, Google and Meta's FAIR research group. It presents its work as "open by default."
The VL release currently has a clearer path to an API than to downloadable weights. Ant Ling's launch materials describe hosted access and a two-week trial through Computrix, where a token plan starts at $20 and is refundable under limited conditions. Ant Ling has not said in the announcement whether Ling-3.0-flash-VL will follow the base model into an open-weight release, or what license would govern it.
That choice will shape the model's reach. Hosted access lets Ant put the model in developers' hands quickly and control the inference environment required for browser and GUI demonstrations. An open-weight release would allow independent researchers to test the claimed score, inspect the architecture and measure whether the visual feedback loop survives outside Ant's stack.
Ling-3.0-flash-VL enters a crowded field that already includes multimodal systems from Google, OpenAI, Anthropic, Alibaba and DeepSeek. Ant's pitch is narrower and practical: take a fast model, give it visual feedback and let it work through interfaces. The launch gives developers enough API surface to test that proposition. The harder proof will come when those agents leave Ant's demonstrations and encounter the messy, shifting software they are supposed to operate.