Published 11 September 2026
Ant Group’s inclusionAI lab published Ling-3.0-flash-VL on 10 September 2026, a model with 124 billion parameters that switches on only 5.5 billion of them per token. It reads images and video as well as text, holds 262,144 tokens of context, and ships under the MIT licence, which permits commercial use. inclusionAI released fp8, fp4 and int4 versions alongside the full-precision weights.
What is Ling-3.0-flash-VL and how is it built?
Ling-3.0-flash-VL is the vision version of Ling-3.0-flash, the text model inclusionAI shipped earlier. It is a mixture-of-experts model, which means it switches on only part of itself for each token. Both versions carry 124 billion parameters in total. The vision model activates 5.5 billion per token, against 5.1 billion for the text one.
Three pieces do the visual work. A ViT encoder turns images and video frames into features, and a small two-layer network maps those features into the same space as the text. VideoRoPE encodes when something happened and not just where it sits in the frame, which is what lets the model answer questions about long video and find a moment inside it. Underneath, a 42-layer backbone alternates two kinds of attention, Kimi Delta Attention and Gated MLA, at a ratio of five to one. That is how it holds a quarter of a million tokens without the memory cost running away.
Thinking mode is on by default, and you switch it off per request.
Ling-3.0-flash-VL benchmarks against Ling-3.0-flash
Read this table as a before-and-after on the same base model rather than a ranking against rivals, because that is the only comparison inclusionAI publishes in numbers.
| What is measured | Ling-3.0-flash-VL | Ling-3.0-flash (text-only predecessor) |
|---|---|---|
| Total parameters | 124 billion | 124 billion |
| Active parameters per token | 5.5 billion | 5.1 billion |
| Inputs accepted | Text, images and video | Text |
| Context window | 262,144 tokens | 262,144 tokens |
| AA Intelligence Index v4.1.1, as quoted by inclusionAI | 42 | 38 |
| Output speed measured by Artificial Analysis | 138.8 tokens per second, against a median of 97 for this size class | Not published in this comparison |
The parameter and context figures come from inclusionAI’s model card. The index score of 42 is inclusionAI’s own quote and refers to version 4.1.1 of the Artificial Analysis Intelligence Index. Artificial Analysis itself now runs version 4.3, on which it lists Ling-3.0-flash-VL at 25, second of 64 open-weights models in its size class against a median of 8. Both numbers are real. They are different tests.
What inclusionAI is not saying about Ling-3.0-flash-VL
The multimodal benchmark comparison exists only as a picture. inclusionAI publishes a chart image on the model card and no table you can read a number out of, so nobody can quote a single MMMU or screen-agent score from this release without rerunning it themselves. For a model whose whole pitch is that vision improves real work, that is the wrong thing to leave out.
The index number in the headline comes from an older scoring version. Quoting 42 on v4.1.1 is not wrong, but Artificial Analysis has moved to v4.3 and scores the model 25 there. A reader who checks the source finds a different number than the one on the model card.
Then there is verbosity, which does not appear on the model card at all. Artificial Analysis measured 160 million output tokens to run its index, against a median of 84 million for comparable models. A reasoning model that thinks at nearly twice the length of its peers is not cheap to serve, whatever the weights cost. inclusionAI also publishes no price of its own. The model is free on OpenRouter at launch, but a launch promotion is not a price.
Licence, hardware and how to run Ling-3.0-flash-VL
- Licence: MIT, commercial use permitted, weights on Hugging Face and ModelScope
- Formats: full BF16 weights plus fp8, fp4 and int4 builds
- Hardware for the full 262,144 token context: four 141GB-class GPUs such as H20-3e or H200, or a four-GPU Blackwell node
- On 80GB cards such as H100 or H800, split it eight ways instead
- Serving: an SGLang docker image ships with the model, and inclusionAI maintains a vLLM fork
- Hosted access: available through Novita, and free on OpenRouter at launch
What this means
This is worth testing now if you build screen or document agents on your own hardware and can free up four large GPUs for them. That is a narrow group, but for them the case is strong. An MIT licence, video input, a quarter-million-token window and 5.5 billion active parameters per token is a combination nobody else is handing out for free this week. A four-person team automating claims handling from scanned documents and screen recordings is the shape of buyer this fits.
It is worth watching rather than adopting if you hoped to swap it into an API and watch a number move. The evidence is thin in exactly the place it matters. There is no readable multimodal benchmark table, the headline index score depends on which version of the index you look at, and the verbosity figure suggests your real bill will be driven by how much the model thinks rather than by what the weights cost. Test it on your own screenshots and your own documents, and count output tokens while you do, because that is where this model will surprise you.
For more information, visit the official announcement of Ling-3.0-flash-VL in the inclusionAI model card.
