Ling-3.2-flash Guide: Multimodal API, Vision, Video

On November 5, 2026, Ant Ling launched Ling-3.2-flash, a native multimodal MoE model with 560 billion total parameters, 25 billion active, and a 1-million-token context. It accepts image, text, and video input over OpenAI- and Anthropic-compatible APIs, with free API access for the first two weeks after launch.

Last updated: November 6, 2026

What shipped on November 5, 2026
Specs compared with earlier Ling models
How to call Ling-3.2-flash
Image and video input limits
Where the model fits
Frequently asked questions
Sources

What shipped on November 5, 2026

Ant Ling’s platform changelog (v2.0.6) lists Ling-3.2-flash as a native multimodal model with 560B total parameters, 25B activated parameters, a 1M context window, and native support for image, text, and video input. The company highlights four focus areas: complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal tasks.

You can try the model in the chat interface and call it through APIs compatible with both OpenAI and Anthropic protocols. Access is free for the first two weeks after launch. The Ant Ling homepage also banners Ling-3.2-flash as live with a limited-time open API offer.

For other recent Chinese multimodal and agent releases covered on FreeAnswers, see Tencent Hy4, ERNIE 5.5 Flash Base, and LongCat-2.5.

Specs compared with earlier Ling models

Ant Ling’s Ling model family page places Ling-3.2-flash next to Ling-3.1-flash and the Ling-3.0 line. Figures below come from that selection guide.

Model	Role	Context	Total / active params	Native multimodal
Ling-3.2-flash	Latest native multimodal (vision, agents, 3D, medical)	1M	560B / 25B	Image, text, video
Ling-3.1-flash	Hybrid reasoning for office, long-horizon coding, agents	1M	560B / 25B	Text-focused peer
Ling-3.0-flash-VL	Earlier VL with visual feedback loop	256K	124B / 5.5B	Image, text, video
Ling-3.0-flash	Cost-efficient high-TPS text model	256K	124B / 5.1B	Text
Ling-3.0-tiny	On-device / offline lightweight	—	7.9B / 1.3B	Text

Ling uses a Mixture-of-Experts design so only a subset of experts activates per token. That is why a 560B-class model can advertise 25B activated parameters while keeping a long context window.

How to call Ling-3.2-flash

Follow Ant Ling’s quickstart. The API base URL is https://api.ant-ling.com/v1, and the recommended model string is Ling-3.2-flash.

Create an Ant Ling account and generate an API key from the prerequisites flow linked in the docs.
Install the OpenAI SDK (pip install openai or npm install openai).
Point the client at base_url="https://api.ant-ling.com/v1" and pass your key.
Call chat.completions.create with model="Ling-3.2-flash" and your messages.
For images or video, send multimodal content blocks as shown in the multimodal tutorial (Base64 image_url / video_url), or use the Anthropic-compatible Messages API documented on the same page.
During the launch window, confirm you are still inside the two-week free period; afterward, check the console pricing and billing pages before production traffic.

Third-party aggregators such as OpenRouter already list earlier InclusionAI Ling models. As of this writing, OpenRouter’s public model list did not yet show a distinct ling-3.2-flash entry, so prefer Ant Ling’s first-party endpoint for the new release.

Image and video input limits

Ant Ling’s multimodal understanding tutorial states that the OpenAI-compatible Chat Completions API and the Anthropic-compatible Messages API both accept text, images, and videos when the model is Ling-3.2-flash. Practical caps from that page:

Images: JPEG or PNG; up to 40 images per request; each image ≤ 16384×784 pixels; each Base64 string ≤ 32 MiB; maximum request body 32 MiB. Inputs are Base64 only.
Videos: MP4, MOV, or WMV; duration ≤ 30 seconds; Base64 ≤ 32 MiB; 1 video per request; maximum request body 32 MiB. Inputs are Base64 only.
Content blocks: use image_url for images and video_url for videos on the OpenAI-compatible path.

Those limits matter for product design. A slide deck of screenshots can fit in one request (under 40 images), but a multi-clip review needs separate calls because video is limited to one file. Keep clips short and compress before Base64 encoding so you stay under 32 MiB.

If you mainly need text coding agents rather than vision, compare with DeepSeek V4.1 Flash or Mistral Large 4. For Google’s latest short-form video generation model, see Gemini Omni 1.2 Flash.

Where the model fits

Ant Ling positions Ling-3.2-flash for workloads that mix perception and multi-step action: reading charts or UI screenshots, driving long multimodal agent loops, 3D-related understanding, and medical multimodal interpretation. The same docs contrast that with Ling-3.1-flash for everyday office and software-agent work, and with Ling-3.0-flash when you need cheaper, high-throughput text.

Because the launch includes a two-week free API window, now is a practical time to benchmark your own image and short-video prompts against your current multimodal stack before paid rates apply. Record token usage in the Ling console so you can estimate cost once the promo ends.

Frequently asked questions

What is Ling-3.2-flash?

Ling-3.2-flash is Ant Ling’s latest-generation native multimodal model in the Ling series. Ant Ling’s November 5, 2026 changelog and model docs describe about 560B total parameters, 25B activated parameters, a 1M-token context window, and native image, text, and video input.

Is Ling-3.2-flash free to use?

Ant Ling’s changelog states the model is free for the first two weeks after launch (announced November 5, 2026). After that window, use Ant Ling’s pricing and billing pages in the developer console for current rates. This guide does not invent post-promo prices.

How do I call Ling-3.2-flash from code?

Use an OpenAI-compatible client pointed at https://api.ant-ling.com/v1 with your API key, and set model to Ling-3.2-flash. Ant Ling’s quickstart shows Python and Node examples. An Anthropic-compatible Messages endpoint is also documented for multimodal requests.

What image and video limits apply?

Per Ant Ling’s multimodal tutorial: images are JPEG or PNG, up to 40 per request, each Base64 string and the request body capped at 32 MiB, with a max image size of 16384×784 pixels. Videos are MP4, MOV, or WMV, duration ≤30 seconds, Base64 ≤32 MiB, and limited to 1 video per request.

How does Ling-3.2-flash differ from Ling-3.1-flash and Ling-3.0-flash-VL?

The model selection guide positions Ling-3.2-flash as the latest native multimodal model for complex visual understanding, multimodal long-horizon agents, 3D interaction, and medical multimodal work. Ling-3.1-flash is a hybrid reasoning text-focused peer with the same 560B/25B and 1M context. Ling-3.0-flash-VL is an earlier multimodal model with 124B total / 5.5B active and a 256K context.

Are the weights open source?

The November 5, 2026 changelog announces API and chat availability with OpenAI and Anthropic protocol compatibility. It does not announce an open-weights release for Ling-3.2-flash. Earlier Ling models have separate open-source notes on Ant Ling’s site; check those pages before assuming weights are available.

Ling-3.2-flash: Power-User Guide to Multimodal API and Limits

Table of contents