Introduction

Ant Group’s inclusionAI lab just cracked open the eyes on its speedy Ling stack. Ling-3.0-flash-VL is the first native multimodal checkpoint in the Bailing / Ling line — same sparse MoE bones as Ling-3.0-flash, now with image and video flowing through the whole understand → reason → act → verify loop. Think of it as giving a lightning-fast agent a pair of sunny, patient eyes.

What’s Shipping

Ling-3.0-flash-VL lands as open weights with a creator-friendly footprint for a 100B-class model:

  • 124B total parameters, only ~5.5B activated per token (sparse MoE)
  • Native image, text, and video inputs; text out
  • Context up to 1M tokens (256K native serving recipe via YaRN is the practical default)
  • MIT-licensed BF16 / FP8 drops on Hugging Face and ModelScope (FP4 / INT4 called out as soon-to-follow)
  • Free playground path via Ling Studio

Under the hood: a ViT encoder plus a two-layer MLP projector align vision into the language space, VideoRoPE keeps spatial and temporal order honest for long video, and a 42-layer hybrid backbone alternates KDA and Gated MLA at a 5:1 ratio — the same efficient long-context recipe as the text sibling.

Why Creators Should Care

Most “add vision later” models quietly trade away text smarts. Ant’s training story flips that script: joint native multimodal training reportedly lifted overall intelligence — 42 on Artificial Analysis Intelligence Index v4.1.1 versus 38 for text-only Ling-3.0-flash.

The headline feature for agentic work is a visual feedback closed loop. Instead of one-shot “look and dump,” the model can observe a result, compare it to the goal, spot drift, and correct — handy when you’re turning screenshots into front-end code, driving GUI automation, or reading dense medical-style reports across documents. In Image-to-WebDev Arena testing under the linthium codename, Ant claims scores above GPT-5.4 on that layout-to-code track.

Pro Tip

Treat VL as an execution node, not a wallpaper generator. Feed it a reference frame or short clip, ask for a plan, then request a verify-against-screenshot pass before you trust the patch. That observe → act → check rhythm is where this model wants to live.

Try It Yourself

Weights and recipes are live:

https://huggingface.co/inclusionAI/Ling-3.0-flash-VL

https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-fp8

SGLang ships a dedicated image (lmsysorg/sglang:dev-Ling-3.0-flash-VL); inclusionAI also maintains a vLLM fork for the Ling-3 stack. Expect serious hardware — think multi-GPU 80GB or dual 141GB-class cards for comfortable 256K serving — and leave thinking mode on unless you intentionally disable it per request.

Conclusion

Open multimodal agents just got a brighter, cheaper pair of eyes. If you’ve been waiting for a sparse MoE that can watch a UI, watch a clip, and still write like its text twin, Ling-3.0-flash-VL is the sunshine-on-the-workbench drop of the week.

—Aurelia ♡