Ling 3.0 Flash VL Guide: Image and Video Input on 5.5B Active

Adding vision to a language model usually means picking a bigger, slower model. Ling 3.0 Flash VL is interesting because it barely moves the cost: it takes Ling 3.0 Flash, a 124B mixture-of-experts model with about 5.1B active parameters, and adds image and video input for roughly 5.5B active. OpenRouter lists it from 10 September 2026, and the MIT-licensed weights are on Hugging Face and ModelScope.

The more surprising claim is that adding vision raised its overall score. inclusionAI reports an Artificial Analysis Intelligence Index v4.1.1 score of 42 for VL against 38 for the text-only base. That is worth understanding before you decide which one to run.

How the vision is attached

The design is conventional at the edges and unusual in the middle. A ViT encoder turns images and video frames into features. A two-layer MLP projects those features into the language model's embedding space. That much is standard.

The backbone is the Ling 3.0 hybrid: 42 layers alternating Kimi Delta Attention and gated MLA at a 5:1 ratio, with sparse mixture-of-experts feed-forward layers. Video timing is handled by VideoRoPE, a positional encoding that records when something happened as well as where it sits in the frame. inclusionAI pitches that at event localisation and long-video questions, the tasks where models that treat video as a bag of frames tend to fall over.

The card groups its claimed abilities as understand, reason and act: reading counts, layouts, charts and documents; doing visual calculation and multi-step reasoning; and interpreting web and software interfaces. The multimodal benchmark scores are in chart images only, so we cannot quote them here.

Why the index score went up

A vision model scoring higher than its own text base on a text-heavy index is not unheard of, but it deserves a sceptical read. inclusionAI does not break down where the gain came from. Plausible contributors are the extra post-training a multimodal release gets, which can lift general reasoning along the way, and a model that simply thinks for longer.

That last point has a cost. Artificial Analysis's own model page, at the time of writing, recorded VL using about 160M output tokens to run its index against a median of 100M for comparable models. A model that thinks longer scores higher and bills more. On the newer index version, v4.3.2, the same page shows a score of 25 and second place of 67 open-weight models in its class; the two versions are not comparable, so do not read 42 to 25 as a decline.

The context window, read carefully

This is the detail most likely to trip someone up. The base Ling 3.0 Flash serves 262,144 tokens. VL's configuration has a native 131,072-token window, and the card's SGLang recipe extends it to 262,144 with YaRN scaling. The OpenRouter listing we synced shows 131,072.

So whether you get 128K or 256K depends on who is serving it and how. Images and video frames consume context quickly, so a long video can hit the lower limit sooner than a text user would expect. Check your endpoint's limit, and sample frames rather than sending everything.

Running it

Self-hosting follows the base model: four 141GB-class GPUs such as the H200 for the full extended context, or eight 80GB cards. There is a dedicated SGLang image, and vLLM needs inclusionAI's fork. The SGLang cookbook covers BF16 and FP8, and inclusionAI also publishes FP8, FP4 and INT4 checkpoints. Thinking is on by default and can be disabled per request with enable_thinking set to false.

Hosted, it is cheap. Our OpenRouter catalogue sync listed it at about $0.06 per million input tokens and $0.18 per million output at the time of writing, roughly three times the text base. Remember that image tokens count as input, so a screenshot-heavy workload costs more than its prompt length suggests.

VL or the text model

• Pick VL when any input is visual: screenshots, scanned documents, charts, UI states, video. Its OpenRouter listing also includes structured outputs, which suits extraction from documents.
• Pick base Ling 3.0 Flash for text-only work at the lowest cost, or when you need a guaranteed 256K text window on a standard endpoint.
• Consider Ling 3.1 Flash for harder text reasoning, but note it is text-only.

Testing it on real visual work

Visual benchmarks transfer poorly, because the images in your product are not the images in a benchmark. Collect twenty real inputs: the dashboards, invoices, error screenshots or clips you would actually send. Write the answer you expect for each. Then run VL with thinking on and off, and score for exact values read from the image, not for a plausible description.

Vision-language models compared lists what to check before committing, and best model for vision tasks covers which visual jobs genuinely need a VLM and which are better served by OCR plus a text model.