Large models can reason, code, and plan, and vision capabilities are no longer scarce. But when it comes to truly entering the real world, every step of the leap—from "seeing" to "getting it right," and then to "getting it right, fast"—is full of challenges.Can a model feed vision back into text after vision is added? Can it reliably deliver professional-scenario outputs based on visual feedback? Can it execute efficiently within visual workflows? These remain questions the industry collectively faces.
With this in mind, we are officially releasing and open-sourcing the first native multimodal large model in the Ling series, Ling-3.0-flash-VL. The model is built by extending the MoE architecture of Ling-3.0-flash, with 124B total parameters and 5.5B parameters activated per inference. It natively supports image, text, and video inputs, with a context window of 256K tokens. Centered on "how to complete real tasks more reliably and efficiently," Ling-3.0-flash-VL focuses on exploring three directions:
- Can the model feed back into text capabilities after vision support is introduced? A common concern when adding vision capabilities to large models is that it will drag down text intelligence. Our training practice yields the opposite conclusion: native multimodal joint training not only expands application boundaries but also boosts text intelligence.
- Can it reliably deliver professional tasks based on visual feedback? In scenarios such as medical report interpretation, front-end development, and GUI automation, the real challenge is not understanding a single image, but continuously acting based on visual information. Ling-3.0-flash-VL introduces a visual feedback mechanism—observing execution results, comparing against goals, detecting deviations, and continuously correcting—turning tasks from one-off "generation" into a closed loop of "observe → act → verify → correct," making execution results more reliable.
- Can the model execute efficiently in visual workflows? Real-world business visual tasks are often not a single "look at the image and describe it" call, but a series of multi-turn processes that require maintaining task state, handling exceptions, and continuously correcting based on visual feedback. Ling-3.0-flash-VL inherits the core advantage of Ling-3.0-flash as an efficient execution node in Agent workflows, balancing output quality and execution efficiency within the visual feedback loop, advancing complete tasks at lower cost and in less time.
01 Native Multimodal Training: Letting Vision and Text Improve Together
For a long time, there has been a widespread concern in the industry: adding vision capabilities to large models will "discount" or even "drag down" their foundational intelligence. In the training of Ling-3.0-flash-VL, we observed a different phenomenon: through native multimodal joint training, the introduction of visual information brought a positive improvement to the model's text capabilities. On the Artificial Analysis Intelligence Index v4.1.1 evaluation, Ling-3.0-flash-VL improved by 4 points over the text-only version (Ling-3.0-Flash); in multimodal comprehensive evaluations, the model also achieved strong performance. This shows that vision is more than just a new input modality—information from images and videos also helps the model build a more complete understanding of concepts, space, relationships, and even the real world.
02 Visual Feedback Loop: Reliably Delivering Professional Tasks
Understanding visual information is only the starting point. Real tasks often require the model to continuously act based on visual information and be accountable for the results of its actions. To this end, Ling-3.0-flash-VL introduces a visual feedback mechanism, advancing tasks from one-off generation to a continuous loop of "observe → act → verify → correct," bringing final results closer to genuinely usable, verifiable deliverables.
Visual Feedback Programming: From Reference Image to Runnable Webpage
Given a webpage screenshot or visual reference, Ling-3.0-flash-VL can understand layout, color scheme, fonts, and component relationships, generate front-end code, and run it in a browser. It can also view the rendering result, compare it with the reference image, and continue modifying the code. This makes webpage replication no longer a one-off "look at the image and generate code," but a development loop driven by visual feedback: visual understanding—layer decoupling—code implementation—rendering comparison—self-correction.
In the official Image-to-WebDev Arena evaluation, Ling-3.0-flash-VL (codename: linthium), with 124B total parameters and only 5.5B parameters activated per inference, achieved competitive performance in the "image-to-webpage" task among multiple mainstream models, scoring higher than GPT-5.4. This result demonstrates the model's potential to achieve high-quality webpage generation at a smaller activated parameter scale, and provides more options for the practical application of visual feedback programming.
🎬 Video demo (seeoriginal WeChat article ）
The front-end replication feature relies on Design Skill for the full experience. We plan to open-source Design Skill soon—stay tuned.
GUI Agent: Continuously Understanding Interfaces, Advancing Cross-Tool Tasks
For cross-tool tasks such as "find data on a webpage, organize it into Excel, then generate a chart," Ling-3.0-flash-VL can identify page structure and application state, break down operation steps, and complete clicks, input, scrolling, and application switching. Through the visual feedback loop of "act → observe → calibrate," it promptly detects deviations and adjusts strategies, continuously advancing tasks within the interface. Compared with generating all operation instructions at once, this mechanism of watching while acting and adjusting while advancing is closer to how humans work in real environments, and provides a foundation for models to enter complex software workflows.
🎬 Video demo (seeoriginal WeChat article ）
AI Health Assistant: Assisted Interpretation of Medical Reports
Medical report interpretation is a high-frequency application scenario for Ant Afu. The model not only needsmedical expertise to accurately identify abnormal indicators and reference ranges; it also needscross-document association capability to integrate multi-page, multi-period data such as blood tests, biochemistry, and imaging into a complete health picture. At the same time, with the "extract → verify → correct" retrospective loop, it repeatedly reconciles values and checks for omissions, ensuring rigorous and reliable interpretation results.
Ling-3.0-flash-VL possesses comprehensive cross-image-text and cross-document understanding capabilities. It can extract key indicators, organize information associations, and visually display health risks with a "red-yellow-green light" system, while also supporting comparative analysis of multiple reports to help users read and organize health information more clearly.
(Note: This capability is for auxiliary reference only and cannot replace professional medical diagnosis.)
🎬 Video demo (seeoriginal WeChat article ）
03 Efficient Visual Node: Lower Cost, Faster Speed
Ling-3.0-flash-VL inherits the ultra-fast inference and flexible deployment advantages of Ling-3.0-flash. The model introduces an arbitrary-resolution vision encoder and VideoRoPE, natively supporting image-text and video multimodal inputs; the language backbone retains the 42-layer hybrid architecture, with KDA and Gated MLA arranged alternately (ratio 5:1). With 124B total parameters, only 5.5B parameters are activated per inference, greatly reducing compute overhead.
Whether it's continuous follow-up questions in real-time video conversations or multi-turn visual interactions across tools, the model maintains low-latency responses. In visual Agent workflows that require repeated "observe—act—verify—correct" cycles, faster single-step reasoning and lower token consumption mean a clear efficiency gain. Ling-3.0-flash-VL delivers a strong experience in the following scenarios:
- Real-time visual interaction: open the camera and chat while watching; the model responds instantly as the scene changes, and the conversation flows naturally without lag;
- Visual Agent workflows: from recognizing the interface and breaking down steps to executing operations and verifying results, multi-turn execution is tight and smooth;
- Long-horizon tasks: native support for 256K tokens easily handles long document parsing, long video analysis, and the historical memory of multi-step Agents.
Real-time visual interaction: instant answers, recognize anything
With a phone camera, Ling-3.0-flash-VL can identify plant and animal species, ingredients, architectural landmarks, and car models, and continuously describe the scene in front of you; for foreign-language signs and text, it can also recognize and translate in real time. These capabilities are already serving users in the Lingguang app—visual understanding is no longer a lab demo, but an information service you can carry in your pocket and use anytime, anywhere. From "seeing" to "getting the answer" in one step.
🎬 Video demo (seeoriginal WeChat article ）
Real-time learning companion: watch and ask, get instant answers
Ling-3.0-flash-VL has strengthened its subject knowledge and visual reasoning capabilities. In courses or educational videos, the model can use the current frame to recognize formulas, diagrams, and whiteboard content, provide context-relevant explanations, and continue answering follow-up questions or generating exercises around the same knowledge point. Video is no longer one-way playback, but an efficient learning space where you can ask questions at any time and get instant responses.
🎬 Video demo (seeoriginal WeChat article ）
Long-horizon tasks: hour-level long-video highlight editing Agent
The difficulty of highlight editing for ball games lies in the fact that goals are often buried in high-speed motion, multi-person occlusion, and fast cuts, so the start and end boundaries are naturally blurry—for example, when the buildup begins and when the celebration ends; single-frame judgment is highly error-prone.
Faced with long match videos, Ling-3.0-flash-VL uses a visual feedback loop of "locate→clip→verify→correct" to fine-tune segment by segment, ultimately producing a smooth highlight reel. Thanks to the low token cost of each step and the native 256K context window memory, even when processing hour-level videos, the model can keep tracking and fine-tuning without missing key clues. From highlight recognition to final delivery, Ling-3.0-flash-VL makes highlight editing more efficient.
🎬 Video demo (seeoriginal WeChat article ）
04 Limitations and Future Plans
Currently, in tasks such as long-tail scenario perception and reasoning, long-chain Agent execution, and long-horizon full-stack coding, Ling-3.0-flash-VL still lags behind the best models; there is also room for improvement in the model's exception recovery and result stability in real environments.
Next, we will focus on three directions:
- Strengthen visual perception and reasoning: improve the ability to understand, reason about, and verify complex charts, mathematical relationships, open-world information, and cross-image content.
- More stable visual Agents: focus on improving long-chain goal retention, error recovery, and end-to-end task completion rates, and enhance the model's execution stability in real workflows.
- Deeper penetration into professional domains: in areas such as front-end development, document intelligence, and medical report interpretation, continue to refine related evaluations, evidence tracing, and safety mechanisms, and promote the gradual public release of evaluation methods.
The Ling-VL series will continue to evolve along the directions of native training, visual feedback, and professional domain penetration, enabling the model to deliver verifiable, reusable results in more complex real-world environments.
Open Intelligence for the Real World.
05 Try Ling-3.0-flash-VL now
Ling-3.0-flash-VL is now available onLing Studio（https://chat.ant-ling.com/chat） and isfree to try.
Download Ling-3.0-flash-VL
Ling-3.0-flash-VL's BF16 and FP8versions are now open source, and the FP4 and INT4 versions will be released soon.
🤗 Hugging Face
- BF16： https://huggingface.co/inclusionAI/Ling-3.0-flash-VL
- FP8： https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-fp8
🤖 ModelScope
- BF16： https://www.modelscope.cn/models/inclusionAI/Ling-3.0-flash-VL
- FP8： https://www.modelscope.cn/models/inclusionAI/Ling-3.0-flash-VL-fp8
Developers are welcome to download and try Ling-3.0-flash-VL and explore more multimodal applications and Agent workflows with us.
Click [Read original article]to visit the Hugging Face page for Ling-3.0-flash-VL.
Model releaseMultimodalPrevious post: Improving the intelligence-efficiency ratio of large models, starting by shortening the wait for one data validationNext post: Open models × open evaluation: Ling-3.0-flash-Fin and FinFIRST officially open source