Launch announced on the company's website
inclusionAI (Ant Ling), the AI unit of China's Ant Group, announced the launch of a new model, "Ling-3.2-flash," in a notice on its website dated Nov. 5. It describes the model as able to read images and video in addition to text and made it available through its developer API. The update comes 2 weeks after the improved Ling-3.1-flash-1019, announced on Oct. 22.
@aitrackerbot, an X account that tracks the appearance of models, had reported the launch on Nov. 5 (Nov. 6 Japan time). No post from the company's official X account has been confirmed. The text of the Chinese-language article linked from the website notice could not be confirmed either.
inclusionAI's Ling 3.2 Flash is out 🚀 It's live on inclusionAI's API platform: 560B total and 25B active parameters, a 1M-token context window, and native image, text and video input. The weights aren't on Hugging Face yet.
Same scale as 3.1, with image and video support
According to the developer documentation, Ling-3.2-flash has 560 billion total parameters, using only about 25 billion per pass. That is the same scale as Ling-3.1-flash. The context window is up to 1 million tokens, it natively accepts image, text and video input, and the company says it can be used to analyze long documents and long videos.
Until now, the Ling model in the series that accepted image and video input was Ling-3.0-flash-VL, a 124 billion-parameter model released in September. Ling-3.2-flash extends that support to the 560 billion-parameter scale. The company cites complex image understanding, long-running agent tasks involving images and video, 3D understanding and manipulation, and medicine as areas of focus. The documentation does not mention audio input, which is handled by the separate Ming series, such as "Ming-3.0."
However, 1 million tokens is the upper limit stated in the documentation, and the length actually usable through the API during the free period has not been confirmed. For Ling-3.1-flash, the API in the early period of availability offered 256,000 tokens, according to orcarouter, an AI model routing service.
| Model | Context window | Total parameters | Parameters used per pass |
|---|---|---|---|
| Ling-3.2-flash | 1 million tokens | 560 billion | 25 billion |
| Ling-3.1-flash | 1 million tokens | 560 billion | 25 billion |
| Ling-3.0-flash-VL | 256,000 tokens | 124 billion | 5.5 billion |
Name also appears in an inference software proposal
In a proposed change submitted on Nov. 5 to the open inference software "vLLM," Ling-3.2-flash appears on the same row as Ling-3.0-flash-VL in the list of supported models. The proposal has not yet been merged. Based on this proposal, @aitrackerbot reported that the model uses the 3.0-flash-VL architecture, but the architectural relationship has not been confirmed in the company's documentation.
inclusionAI's Ling 3.2 Flash is on the way 👀 A new vLLM PR from a Ling contributor adds inclusionAI/Ling-3.2-flash to the supported models list. It runs on the Ling 3.0 Flash VL architecture and takes text, images and video, and the PR adds video and YaRN long-context support. The Hugging Face repo isn't public yet.
No performance figures, and no third-party evaluations yet
The developer documentation does not include benchmark figures for Ling-3.2-flash. Artificial Analysis, which compares AI models under the same conditions, also had no page for the model as of Nov. 6. For the previous version, Ling-3.1-flash-1019, the company announced a score of 44 on Artificial Analysis's intelligence index, and how much the new model has improved cannot be confirmed.
Free for 2 weeks; later pricing not in the price list
According to the price list, Ling-3.2-flash can be used free of charge for 2 weeks after release. The company says prices will return to the regular rate when the discount period ends, but the price list has no row for Ling-3.2-flash and does not show the unit price after it becomes paid or the end date of the free period.
A reference point is the pricing for Ling-3.1-flash, a model of the same scale. Its list price on OpenRouter and ZenMux is $0.30 for input and $0.90 for output per 1 million tokens, discounted for a limited time to $0.075 for input and $0.225 for output. On the company's own API, the list price is 2 yuan for input and 6 yuan for output, and 0.50 yuan for input and 1.50 yuan for output after the discount. "M3.1 Flash" from China's MiniMax, which has a similar context window and input types, is offered at $0.15 for input and $0.60 for output.
- Ling-3.1-flash (list)
- Ling-3.1-flash (discounted)
- MiniMax M3.1 Flash
Weights not released, and no release plan given
The weights (trained parameters) for Ling-3.2-flash had not been released on Hugging Face as of Nov. 6. The name does not appear in the company's model list on Hugging Face, and the documentation gives no release plan.
The company has so far released the weights of its main models under the MIT License, which also allows commercial use. Ling-3.1-flash-1019's weights were released on Oct. 27, 5 days after it was announced, and the weights for Ling-3.0-flash-VL have also been released. MiniMax M3.1 Flash released its weights 4 days after its API launch. For companies that want to run the model in their own environments, the focus is whether the same pattern will follow this time.
A string of Ling updates since September
Since September, the company has released Ling models in quick succession. On Nov. 5, @aitrackerbot's report and the notice on the company's website carried the same date.
- Ling-3.0-flash-VL, with image and video support, releasedOfficial
- Ling-3.1-flash announcedOfficial
- Improved Ling-3.1-flash-1019 announcedOfficial
- @aitrackerbot reports the name in a vLLM proposal and the start of API availabilityLeak
- Company website announces launch of Ling-3.2-flashOfficial
