InclusionAI, linked to Ant Group, releases Ling and Ring 2.6 models — including 1T checkpoints

InclusionAI, a team publicly tied to Ant Group, has published a technical report for its Ling-2.6 and Ring-2.6 artificial intelligence model families and is making the model checkpoints publicly downloadable on Hugging Face under an MIT license, including two trillion-parameter variants.

The report, titled “Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale,” was submitted to arXiv on June 13 in the computer science and language category, cs.CL. Ang Li is listed first among the authors, and Yuxin Tian is named as the submitting contact. The paper’s abstract says, “We open-source all checkpoints in the 2.6 family to support further research and development in practical agentic intelligence.”

That combination — a formal technical report paired with openly accessible weights for very large models — is the notable part of this release. Open-weight AI models are not new, and Ant Group has been releasing models in this line for months, but it is still relatively uncommon to see trillion-parameter-class systems distributed publicly alongside a detailed report describing their design and training approach.

The public Hugging Face pages under InclusionAI currently include Ling-2.6-flash, Ling-2.6-1T and Ring-2.6-1T. The model cards describe them as official open-source releases and list the license as “mit.” One card says, “Today, we announce the official open-source release of Ling-2.6-flash.”

According to the model pages, Ling-2.6-flash has about 104 billion total parameters, with about 7.4 billion active parameters. Ling-2.6-1T and Ring-2.6-1T are each listed at about 1 trillion total parameters, with about 63 billion active parameters. The same pages advertise long context windows in roughly the 256,000- to 262,000-token range and include deployment instructions for common inference tools such as Transformers, vLLM and SGLang, as well as quantized versions.

In the report’s abstract, InclusionAI describes Ling-2.6 as optimized for low-latency, or fast-response, generation, while Ring-2.6 is positioned for deeper reasoning and more advanced agent-style workflows, meaning systems that can carry out multistep tasks. The team says it upgraded a Ling-2.0 base model rather than training the new family from scratch, using what it calls architectural migration pre-training and large-scale post-training. The report also describes a hybrid linear attention design combining Lightning Attention with MLA, and an agent-training reinforcement learning framework called KPop for Ring-2.6-1T.

The trillion-parameter figure is plausible in part because these are mixture-of-experts models, a design that allows a model to contain a very large total number of parameters while activating only a smaller subset for each token. That is why the vendor can cite around 1 trillion total parameters while reporting much lower active parameter counts.

The release also fits a broader pattern. Public materials tie InclusionAI to Ant Group and its Ant Ling, or Bailing, model effort, and the group previously open-sourced Ling/Ring 2.5 trillion-parameter variants in February 2026. In that sense, the 2.6 launch looks less like a one-off disclosure and more like the next step in an ongoing open-weight strategy.

What is directly verifiable here is the publication of the arXiv report and the public availability of the model weights on Hugging Face under a permissive MIT license. The model cards also include benchmark and “open-source SOTA” claims, but those should be treated as vendor-reported. Some results are labeled as internal evaluations or tied to specific leaderboard snapshots, and the report does not provide independent, peer-reviewed validation of those performance claims.