Ring-2.5-1T

2026-02-10 · Publisher : Ant Group · Domain : Artificial Intelligence

Ant Group open-sources Ring-2.5-1T, a 1-trillion-parameter reasoning model with hybrid linear attention, under the MIT license, which reaches gold-medal level at IMO 2025 and CMO 2025 according to the vendor.

### Context Ring-2.5-1T succeeds Ring-1T (October 2025). In the meantime, the race has shifted toward long-running agents, where the cost of generation over long contexts becomes the bottleneck. The inclusionAI team responds with a hybrid architecture, Ling 2.5, derived from Ling 2.0 through incremental training. ### What the model brings - Hybrid linear attention at scale: the grouped attention layers of Ling 2.0 are converted into a mix of 1 MLA layer for 7 linear attention layers (Lightning Linear), which, according to the vendor, reduces memory access by more than 10 times and multiplies generation throughput by more than 3 beyond 32,000 tokens. - More rigorous reasoning: addition of dense rewards that evaluate the rigor of reasoning in addition to final correctness; at IMO 2025 (maximum score 42), 35 points, gold-medal level; at CMO 2025 (maximum score 126), 105 points, above the gold threshold (78) and the national team selection bar (87). These exams are self-tested. - Long task execution: fully asynchronous agent reinforcement learning, with adaptation to agent programming frameworks such as Claude Code and OpenClaw according to the model card. - Intensive thinking mode ("heavy thinking"): parallel reasoning and synthesis to trade compute for quality at inference time. - Detailed IMO and CMO solutions published on GitHub. ### Architecture and specifications - Architecture: 1-trillion-parameter MoE, about 63 billion active parameters (versus 51 billion for Ling 2.0), hybrid MLA and Lightning Linear attention at a 1:7 ratio. - Context: 128,000 tokens extended to 256,000 via YaRN. Modalities: text. - License: MIT. Weights: Ring-2.5-1T on Hugging Face and ModelScope (BF16 and FP8 supported by SGLang). - Training: token count and knowledge cutoff not disclosed. ### Cost and access The model card states that the online trial and the API are coming soon: no pricing is published. For self-hosting, the model card's example reuses a multi-node configuration (4 nodes, tensor parallelism 8 and pipeline parallelism 4) with a dedicated SGLang branch. | Item | Information | | --- | --- | | Official API | Pricing not disclosed (announced "soon") | | Self-hosting | Multi-node (example: 4 nodes), dedicated SGLang branch | | License | MIT | ### Benchmarks | Metric | Value published by the vendor | | --- | --- | | IMO 2025 (maximum score 42) | 35 points, gold-medal level (self-tested) | | CMO 2025 (maximum score 126) | 105 points; gold threshold 78, national bar 87 (self-tested) | | Generation throughput beyond 32,000 tokens | More than 3 times that of the previous version (announcement) | | Memory access beyond 32,000 tokens | More than 10 times lower (announcement) | Source: Ring-2.5-1T Hugging Face model card. The comparison table with closed models is an image with no values in text; no numerical comparison with closed models is therefore included here. ### Versus closed models The model card compares Ring-2.5-1T with DeepSeek-V3.2 and Kimi-K2.5 (open) and with GPT-5.2, Gemini 3 Pro and Claude Opus 4.5 (closed) and claims state-of-the-art level among open models on IMO AnswerBench, AIME 26, HMMT 25, LiveCodeBench, ARC-AGI-V2, Gaia2, Tau2-bench and SWE-Bench Verified, but the values appear only in an image and cannot be reproduced here. The vendor does not claim parity with closed models in the text. The olympiad results are self-tested, with the solutions published, which allows verification. No pricing is available to gauge the cost gap. What is documented is the efficiency advantage of hybrid attention on long contexts, relevant for long-running agents hosted in-house. ### Key takeaways Ring-2.5-1T is the first 1-trillion-parameter reasoning model with hybrid linear attention, with an announced throughput gain of more than 3 times on long contexts. Numerical comparisons with closed models are not available in text: the images in the model card must be read to verify them.

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