Ant Group has introduced and made public the Ling-3.0-flash-VL model, which is the first native multi-modal model in the Ling series. The release was conducted through the organization inclusionAI on the Hugging Face and ModelScope platforms.
The model is built upon the Ling-3.0-flash mixture-of-experts backbone and has a total size of 124 billion parameters, with approximately 5.5 billion parameters activated per token. It can accept images, text, and video as input, and the company specifies a context window of 256 thousand tokens for processing long documents and video materials. Weights in BF16 and FP8 formats were published on September 9th, with versions in FP4 and INT4 formats planned soon.
A key design decision was the implementation of a vision-based feedback loop. This loop treats perception not as a one-time step in generating a caption, but as an integral part of a continuous workflow. The model is instructed to observe, act, verify, and correct its results based on visually displayed or generated data.
Ant Group positions this model for automating graphical user interfaces (GUIs), frontend code generation, and medical report analysis. In tests related to converting images to web pages, the model demonstrated the ability to restore relationships between components and layouts, after which it reviewed the generated code by comparing it with rendering results. Under the pseudonym linthium in the Image-to-WebDev Arena, Ant Group reported that its score surpassed that of GPT-5.4.
As a GUI agent, the model can parse interface structure and link actions between different tools. In clinical report scenarios, it is designed to consolidate information from various documents and identify risks, rather than just summarizing the content of a single page.
According to Ant Group's statements, joint training using native multi-modal data improved both the model's textual capabilities and visual skills. According to the Artificial Analysis Intelligence Index v4.1.1, Ling-3.0-flash-VL received a score of 42, which is four points higher than the text-only version Ling-3.0-flash.
Hugging Face materials structure the evaluation across three dimensions: understanding, reasoning, and action. These dimensions include working with complex layouts and documents, multi-step evidence-based verification, and executing interface-driven tasks. This approach aligns with the company's strategic goal that open multi-modal weights should help agents close loops in software and clinical workflows, rather than just answering static visual questions.
The model architecture combines an arbitrary resolution vision encoder and a two-layer projector with VideoRoPE temporal encoding, followed by a 42-layer hybrid language backbone. This backbone alternates between KDA and gated MLA blocks in a 5:1 ratio. Thanks to sparse MoE, the computational load remains low even during long visual chat sessions and agent action histories.
Demos are available on Ling Studio. Along with the weights, recipes for SGLang and a fork of vLLM configured for Ling are documented, including parsers for reasoning and tool calling adapted to the Ling-3 chat template.
For developers, this open release represents not just another caption generator, but an early open multi-modal Ling stack that integrates vision into planning and verification processes for agent tasks. It is also the first open Ling release to include a native multi-modal path with an explicit vision feedback loop, rather than just an attached package of visual features.