InclusionAI releases Ring‑2.6‑1T

InclusionAI, the open-source AI organization linked to Ant Group, has released Ring-2.6-1T, a trillion-parameter reasoning model published under the MIT license. The model appeared on Hugging Face about three days before May 17, according to the repository page, which describes it as a flagship system for “real-world complex task scenarios.” The listing gives developers direct instructions for loading the model with Transformers and serving it through vLLM, SGLang and Docker-based tools. (huggingface.co) The Hugging Face files page lists the package at 1.04 terabytes. ### Where did this model actually show up? Hugging Face hosts the public Ring-2.6-1T repository under the inclusionAI organization, with the model card marked “License: mit.” The page says the model is intended for developers, researchers and enterprises to validate, adapt and further develop. (huggingface.co) GitHub and Hugging Face pages show InclusionAI as Ant Group’s open-source AI arm. The InclusionAI GitHub organization says it contains open-source projects from Ant Group focused on AGI-related models, training systems and inference frameworks. ### What is Ring-2.6-1T claiming to be? (huggingface.co) The model card describes Ring-2.6-1T as “a trillion-parameter flagship reasoning model.” The same page says the goal is not only larger scale, but support for agent workflows, engineering development, scientific research analysis, complex business systems and enterprise automation. (huggingface.co) Ant Group said in an October 2025 press release that its Ling model family includes three main series: Ling for non-thinking models, Ring for thinking models and Ming for multimodal models. That earlier release also said Ant had open-sourced Ling-1T and had previously released Ring-1T-preview in September 2025. (github.com) ### What does the MIT license change for users? The Hugging Face listing labels Ring-2.6-1T with the MIT license, one of the most permissive standard software licenses used in open-source projects. The repository also includes direct examples for loading the model through open tooling rather than restricting access to a hosted API. (huggingface.co) The practical result, based on the published instructions, is that users can download weights and run the model through local or self-managed infrastructure if they have the hardware. The model card includes commands for `transformers`, `vllm serve`, `sglang.launch_server` and Docker model runners. (antgroup.com) ### Can most people really run a 1T model locally? The files page lists Ring-2.6-1T at 1.04 TB, which indicates substantial storage and compute requirements even before deployment overhead. The public instructions show ways to serve the model, but the repository page shown in search results does not provide a consumer-grade hardware profile or a minimum GPU count. (huggingface.co) InclusionAI’s earlier open models used sparse mixture-of-experts designs with far fewer activated parameters than total parameters. Earlier InclusionAI documentation for Ling-Lite and Ling-Plus, for example, described 16.8 billion total parameters with 2.75 billion activated, and 290 billion total with 28.8 billion activated. (huggingface.co) Ring and Ling documentation also describe the family as MoE-based. Ring-2.6-1T’s public card, however, does not spell out an activated-parameter figure in the snippets available here. ### How does this fit into Ant Group’s broader model push? Ant Group said in October 2025 that Ling-1T expanded a broader family of open models and called the Ring line its “thinking” series. (huggingface.co) Chief Technology Officer He Zhengyu said then that the open-source release of Ling-1T and Ring-1T-preview was part of Ant’s commitment to collaborative development. The InclusionAI organization page now shows both Ling-2.6-1T and Ring-2.6-1T among its recently updated models, suggesting the new reasoning release is part of a wider 2.6 generation rollout rather than a one-off upload. The same page lists the Ling-2.6-1T model as having appeared 13 days earlier. (huggingface.co) ### What should developers watch next? The Ring-2.6-1T repository was updated within hours of this report, according to the Hugging Face files page, and its eval-results directory was also recently modified. Developers looking for benchmark details, deployment changes or additional documentation will likely find the next updates first in InclusionAI’s Hugging Face repository and GitHub organization pages. (antgroup.com) (huggingface.co 1) (huggingface.co 2)