InclusionAI

According to the official site, inclusionAI is Ant Group's artificial general intelligence initiative, which publishes the Ling, Ring and Ming model families with open weights.

The organization's Hugging Face page describes the family as the Ant Ling AI Model Family. The Ling series groups large language models, the Ring series groups reasoning models and the Ming series groups multimodal any-to-any models. The weights are on Hugging Face and ModelScope.

Among the models cited on the official pages are Ling-1T, with 1 trillion parameters, and Ring-2.5-1T, which the organization itself describes as the first open-source trillion-parameter reasoning model based on hybrid linear attention. The organization also maintains LLaDA2.X, a series of diffusion language models, and audio models in the Ming line.

inclusionAI works as a research lab that publishes models, not as a software product with its own interface.

inclusionAI is for researchers and machine learning teams that want to download open weights of Ant Group's language, reasoning and multimodal models and run them on their own infrastructure.

Teams researching long-horizon reasoning find the Ring series. Teams building text assistants and agents find the Ling series. Teams working with input and output in several modalities, including audio, find the Ming series. Researchers of alternative architectures find LLaDA2.X, with diffusion language models.

The prerequisites are practical. The weights are distributed on Hugging Face and ModelScope, and the sources consulted do not describe an API hosted by inclusionAI itself, so the team has to operate inference. The largest models have 1 trillion total parameters, and the organization also publishes smaller models, which are the simplest starting point for tests.

Each repository has its own license, which should be read before use in a product.

The main features of inclusionAI are the Ling, Ring and Ming families, plus research models such as LLaDA2.X, all with published weights.

• Ling: large language models. Ling-1T has 1 trillion parameters, and an Ant Group announcement about a later version cites a context of up to 1 million tokens.
• Ring: reasoning models. Ring-2.5-1T is described on the model page as the first open-source trillion-parameter reasoning model based on hybrid linear attention.
• Ming: any-to-any multimodal models, with omni variants and audio models listed on Hugging Face.
• LLaDA2.X: a series of diffusion language models, maintained by the inclusionAI team at Ant Group, according to GitHub.
• Distribution: weights on Hugging Face and ModelScope, with code repositories on GitHub.

The organization's Hugging Face page lists more than 200 models in one snapshot, so new variants appear often. Benchmark results published by Ant Group are self-reported and are not included here.

inclusionAI is worth it for teams that want open weights of large, reasoning and multimodal models and have a team to run inference, and less so for teams looking for a ready-made API.

Pros:

• Three families with distinct purposes, language, reasoning and multimodal, published in the same organization.
• Weights on Hugging Face and ModelScope, with code repositories on GitHub.
• The largest models, with 1 trillion total parameters, published with open weights.
• Research architectures beyond the standard, such as the diffusion language models of LLaDA2.X.

Cons:

• The sources consulted describe distribution of weights, not an API hosted by inclusionAI, so running inference is up to the team.
• The listing has more than 200 models from different generations, and you have to choose among families and variants before starting.
• Benchmark results published by Ant Group are self-reported, and the team needs to validate the model with its own data.
• Each repository may have its own license, and checking it is up to whoever will use it in a product.

The alternatives to inclusionAI are other labs that publish weights for large language models.

• DeepSeek: a lab that publishes open weights of language and reasoning models and offers an API.
• Alibaba Qwen: the Qwen family, with several open-weight models in different sizes, plus API access.
• Kimi (Moonshot AI): the Kimi K2 series, a mixture-of-experts model with 1 trillion total parameters, with weights on Hugging Face and an API compatible with the OpenAI interface.
• Xiaomi MiMo: a family with weights on GitHub and Hugging Face and its own API platform.
• Meta Llama: the Llama family, published with open weights under Meta's own license.

The central difference from inclusionAI is the split by purpose. inclusionAI separates language, reasoning and multimodal into three families, and the alternatives tend to combine those uses in fewer model lines. Teams that need a hosted API have, in Kimi and Xiaomi MiMo, platforms of their own described in the sources consulted.