Ring-1T

Ring-1T is the model that opened the trillion-parameter reasoning line of InclusionAI, the open-source arm of Ant Group. The weights went up on Hugging Face on 10 October 2025, a fortnight after a preview version, and the technical report followed on 21 October. It is built on the Ling-1T base model this catalogue also covers, using the same Ling 2.0 architecture: 80 layers, 256 routed experts with 8 activated plus one shared, four dense layers at the front, hidden size 8192, vocabulary 157,184. The published index counts 999.7 billion parameters with 50 billion activated per token, and the context window is 65,536 tokens natively, doubling to 131,072 with YaRN. The main repository ships in BF16, with a separate FP8 repository alongside it. What sets the release apart is the way the maker reported its own testing. For IMO 2025 it ran the model inside its AWorld multi-agent framework and published the result problem by problem: solved 1, 3, 4 and 5 on the first attempt, which is silver medal level; produced a nearly complete geometry proof for problem 2 only on the third attempt; and on problem 6, which no AI entrant solved that year, converged on the answer 4048 when the correct answer is 2112 — the same wrong answer Gemini 2.5 Pro reached. Publishing the wrong answer, and naming the competitor that made the same mistake, is unusual in this segment. The reasoning traces were released in the AWorld repository, so the claim can be inspected rather than taken on trust. At the ICPC World Finals 2025, again the maker's own test, Ring-1T solved 5 problems in three attempts against 6 for GPT-5-Thinking and 3 for Gemini 2.5 Pro. The engineering contribution behind the model is the training method rather than the architecture. Reinforcement learning on a mixture-of-experts model of this size breaks down because the training engine and the inference engine compute operators slightly differently, and that gap grows with sequence length and training steps; the maker's charts show the standard GRPO algorithm collapsing after relatively few steps. Its answer, called Icepop, corrects the distributions with masked two-sided truncation and keeps the gap flat. The supporting infrastructure — a training system called ASystem with a shared memory pool for training and inference, and a sandbox reward service handling ten thousand requests a second across more than ten programming languages — is described in the card, and the AReaL framework from that stack was open-sourced. As with every release from this publisher, the licence claim is narrower than the label. The repository metadata says MIT, but the model card licenses the code repository under MIT and links to a file in a separate GitHub project; among the 170 files of the weights repository there is no licence document. Unlike the later Ring-2.5-1T and Ring-2.6-1T, this model was available outside China from day one: the card points overseas developers to ZenMux for chat and API access, and Chinese users to the maker's own Ling Chat. Downloads have since fallen to a few hundred a month, well below the 2026 successor that most self-hosting users settled on.

▸Videos

No videos yet.

▸Reviews