Ling-1T

2025-10-09 · Publisher : Ant Group · Domain : Artificial Intelligence

Ant Group, through its inclusionAI team, opens Ling-1T, a language model with 1 trillion parameters (about 50 billion active), non-reasoning, under the MIT license, trained in FP8 precision.

Ant Group, through its inclusionAI team, opens Ling-1T, a language model with 1 trillion parameters (about 50 billion active), non-reasoning, under the MIT license, trained in FP8 precision. ### Context In early October 2025, the largest open models are Kimi-K2 (about 1.043 trillion parameters) and DeepSeek-V3.1 (671 billion). Ant Group had published the small models Ling-mini-2.0 (16 billion) and Ling-flash-2.0 (103 billion) in September; Ling-1T is the largest of the Ling 2.0 family and serves as the base for the reasoning model Ring-1T. ### What the model brings - High sparsity: 1/32 of experts activated, about 51 billion active parameters out of 1 trillion, guided by the team's own scaling laws (Ling Scaling Laws). - Largest foundation model known to be trained in FP8 according to the vendor: about 15% end-to-end speed gain and a loss gap of 0.1% or less compared with BF16 over 1 trillion tokens. - Evolutionary chain of thought (Evo-CoT) and sentence-level policy optimization (LPO): each sentence is treated as an action unit, between the token level and the sequence level. - Hybrid syntax, function and aesthetics reward for front-end code generation: Ling-1T ranks first among open models on ArtifactsBench according to the model card. - Tools without massive agent data: about 70% tool-calling accuracy on BFCL V3 with a simple light tuning (vendor claim). ### Architecture and specifications - Architecture: MoE with 1 trillion parameters, 51 billion active (technical report), multi-token prediction layer, auxiliary-loss-free routing with sigmoid scoring, QK normalization. - Training: more than 20 trillion tokens, including more than 40% reasoning-rich data at the end of pre-training. - Context: 32,000 tokens extended to 128,000 by YaRN. Modalities: text. - License: MIT. Weights: Ling-1T on Hugging Face and ModelScope. Knowledge cutoff: not disclosed. ### Cost and access No official API pricing was found for Ling-1T: the model card points to the third-party platform ZenMux. The self-hosting hardware is not specified; the trillion-parameter scale implies several GPU nodes. | Item | Information | | --- | --- | | Official API | Pricing not disclosed | | Self-hosting | Not disclosed; multi-node deployment expected at this size | | License | MIT | ### Benchmarks | Benchmark | Ling-1T | GPT-5-main | Gemini 2.5 Pro (lowthink) | Kimi-K2-Instruct-0905 (open) | | --- | --- | --- | --- | --- | | AIME 2025 (Pass@1) | 70.42 | 59.43 | 70.10 | 50.16 | | HMMT 2025 (Pass@1) | 47.08 | 36.98 | 60.73 | 38.80 | | LiveCodeBench (Pass@1) | 61.68 | 48.57 | 45.43 | 48.95 | | CodeForces (rating) | 1901 | 1120 | 1675 | 1574 | | ArtifactsBench | 59.31 | 41.04 | 60.28 | 44.87 | | ARC-AGI-1 (Pass@1) | 43.81 | 14.06 | 18.94 | 22.19 | | GPQA-Diamond (Pass@1) | 72.98 | 71.31 | 71.81 | 73.93 | | MMLU-Pro (Acc.) | 82.04 | 81.94 | 82.13 | 81.03 | | BFCL-V3 (function calling) | 69.64 | 50.27 | 63.31 | 71.05 | Source: Ling 2.0 technical report (arXiv 2510.22115), table 8. Scores measured by the Ling team (vendor); GPT-5-main and Gemini 2.5 Pro are evaluated with a low reasoning effort. ### Versus closed models In the report's table, Ling-1T is ahead of GPT-5-main and Gemini 2.5 Pro (low effort) on code (LiveCodeBench 61.68 versus 48.57 and 45.43; CodeForces 1901 versus 1120 and 1675) and on AIME 2025 against GPT-5-main (70.42 versus 59.43), and it is tied with Gemini 2.5 Pro on that test (70.10). It trails Gemini 2.5 Pro on HMMT 2025 (47.08 versus 60.73) and is at about the same level on general knowledge (MMLU-Pro 82.04, GPQA-Diamond 72.98). The comparison covers low reasoning effort configurations: it says nothing about the high-reasoning versions of these models, which are stronger. The price gap is not established for lack of public pricing. For the enterprise, the interest is that a non-reasoning model of this size, under the MIT license, sits at the level of closed references in economy mode, but running it requires multi-node infrastructure. ### Key takeaways Ling-1T shows that an open MoE with 1 trillion parameters can be trained in FP8 with good stability and hold its own against closed models in low reasoning effort mode. Above all, it served as the base for Ring-1T, published five days later.

Tags : LLM, Open Source, inclusionAI, Ant Group, Ling, MIT, MoE, 1T