Ring-1T

2025-10-14 · Publisher : Ant Group · Domain : Artificial Intelligence

Ant Group opens Ring-1T, a reasoning model with 1 trillion parameters (50 billion active) under the MIT license, which reaches silver-medal level at the 2025 International Mathematical Olympiad in natural language.

### Context Ring-1T is the reasoning counterpart of Ling-1T, published five days earlier: it is trained on the same base, Ling-1T-base. A preview had been opened on September 30, 2025. To position itself, the vendor compares Ring-1T with GPT-5-Thinking (High) and Gemini 2.5 Pro on the closed side, and with DeepSeek-V3.1-Terminus and Qwen3-235B-A22B-Thinking-2507 on the open side. ### What the model brings - Stable reinforcement learning at scale: the Icepop algorithm corrects the gap between training engine and inference engine specific to MoEs, which worsens with training duration in GRPO. - ASystem, an in-house reinforcement learning system for the trillion-parameter scale (single-controller and SPMD architecture, serverless sandbox starting in milliseconds); the AReaL framework is open. - Natural-language reasoning: Ring-1T solves problems 1, 3, 4 and 5 of IMO 2025 in one attempt (silver-medal level) and produces on a third attempt an almost complete proof of problem 2, according to the model card. - Competitive programming: at the ICPC 2025 world finals, with three attempts, Ring-1T solves 5 problems against 6 for GPT-5-Thinking and 3 for Gemini 2.5 Pro (vendor test, ongoing). - FP8 version published alongside the BF16 version. ### Architecture and specifications - Architecture: Ling 2.0 MoE, 1 trillion parameters in total, 50 billion active; Ling-1T-base base. - Context: 64,000 tokens extended to 128,000 by YaRN. Modalities: text. - Post-training: large-scale RLVR (verifiable rewards) followed by RLHF to rebalance general capabilities. - License: MIT. Weights: Ring-1T and Ring-1T-FP8 on Hugging Face and ModelScope. Knowledge cutoff: not disclosed. ### Cost and access No official API pricing is published in the sources consulted; the model card points to the Ling Chat interface and the third-party platform ZenMux. For self-hosting, the model card example spreads the model across 4 nodes (tensor parallelism 8 and pipeline parallelism 4). | Item | Information | | --- | --- | | Official API | Pricing not disclosed | | Self-hosting | SGLang example on 4 nodes (tp 8, pp 4), BF16 and FP8 versions | | License | MIT | ### Benchmarks | Benchmark | Ring-1T | GPT-5-Thinking (High) | Gemini 2.5 Pro | DeepSeek-V3.1-Terminus-Thinking (open) | | --- | --- | --- | --- | --- | | AIME 2025 (Avg@64) | 93.40 | 94.60 | 88.00 | 89.06 | | HMMT 2025 (Avg@16) | 86.72 | 93.30 | 82.50 | 86.10 | | LiveCodeBench v6 (Avg@4) | 78.30 | 80.60 | 70.65 | 75.33 | | CodeForces (rating) | 2088 | 1918 | 1837 | 2073 | | ARC-AGI-1 | 55.94 | 65.70 | 45.44 | 40.62 | | Arena-Hard v2 (win rate) | 81.59 | 82.91 | 79.20 | 60.27 | | HealthBench | 57.93 | 67.20 | 49.39 | 50.19 | | GPQA-Diamond | 78.63 | 86.05 | 86.40 † | 81.00 † | | Humanity's Last Exam | 16.03 | 24.84 | 21.60 † | 17.82 | | BFCL v3 | 68.82 | 57.21 | 61.36 | 62.01 | Source: technical report arXiv 2510.18855, table 1. Scores measured by the authors (vendor); †: value taken from the official documentation of the model concerned. ### Versus closed models Ring-1T is close to GPT-5-Thinking (High) on several reasoning tests: 93.40 versus 94.60 on AIME 2025, 78.30 versus 80.60 on LiveCodeBench v6 and 81.59 versus 82.91 on Arena-Hard v2. It is ahead of Gemini 2.5 Pro on AIME 2025 (88.00), on LiveCodeBench (70.65), on CodeForces (2088 versus 1837) and on ARC-AGI-1 (55.94 versus 45.44). The gap widens against GPT-5-Thinking on HMMT 2025 (86.72 versus 93.30), ARC-AGI-1 (55.94 versus 65.70), HealthBench (57.93 versus 67.20) and Humanity's Last Exam (16.03 versus 24.84), and the model is far behind in code editing on Aider (78.57 versus 95.49 for GPT-5-Thinking). No pricing is published, so the price gap cannot be quantified. For the open or closed choice, Ring-1T offers an MIT license and full weights, but running it requires a multi-node infrastructure. ### Key takeaways Ring-1T is the first open reasoning model with 1 trillion parameters, with vendor results close to GPT-5-Thinking on mathematics and code. It remains behind on knowledge and specialized tasks, and running it implies a heavy infrastructure.

Tags : LLM, Open Source, inclusionAI, Ant Group, Ring, MIT, MoE, Raisonnement