Ring-2.5-1T

Ring-2.5-1T is the reasoning model that InclusionAI, the open-source arm of Ant Group, published on Hugging Face on 10 February 2026, with the model card revised five days later. The maker presents it as the first open trillion-parameter thinking model built on a hybrid linear attention architecture. It sits between two models this catalogue already covers: the Ling-1T base flagship that opened the trillion-parameter line, and Ring-2.6-1T, the newer reasoning flagship from May 2026. The architectural change is the point of the release. The publisher took the Ling 2.0 architecture and, through incremental training rather than a fresh run, converted its grouped-query attention into a 1:7 mix of multi-head latent attention and Lightning Linear attention. The stated purpose is throughput during long chains of reasoning: the maker reports over ten times less memory access overhead and more than three times the generation throughput on sequences beyond 32,000 tokens, measured on eight H20-3e and eight H200 GPUs at batch size 64. The cost of the change is a heavier forward pass — activated parameters rose from 51 billion in Ring-1T to 63 billion here — and the maker's argument is that the linear attention mix more than pays that back the longer the answer runs. The headline results are mathematical and self-reported. On IMO 2025 the model scored 35 points out of 42, which is gold medal level; on CMO 2025 it scored 105 out of 126, against a gold threshold of 78 and a national squad cutoff of 87. Both figures are the maker's own tests, and the full solutions were published in the project repository, which is more than most publishers of such claims offer. Beyond that, the card contains no results table at all — the comparisons against DeepSeek-v3.2-Thinking, Kimi-K2.5-Thinking, GPT-5.2, Gemini 3.0 Pro and Claude Opus 4.5 exist only as chart images, so no individual score can be quoted from them. This is the third model of this publisher in the catalogue with that same problem. The licence needs stating carefully. The repository metadata carries an MIT label, but the sentence in the model card is narrower than the label: it says the code repository is licensed under MIT and links to a file in a separate GitHub project. Among the 171 files of the weights repository there is no licence document. The same arrangement recurs across this publisher's releases. One number puts the release in perspective. In the 30 days to 29 August 2026 Ring-2.5-1T was downloaded 9,054 times, while its own successor Ring-2.6-1T — a model of almost identical size, released three months later — was downloaded fewer than 500 times in the same window. Neither model is resold by any provider on OpenRouter, and at the time of the release the maker's own chat and API were still marked as coming soon. Whoever runs this model runs it themselves, and mostly still runs this version rather than the newer one. The maker itself names token efficiency and instruction following as the weak points of the release.

▸Videos

No videos yet.

▸Reviews