Atlas

Models

← All models

Ring-1T

InclusionAI · 2025-10-14 · 999.7B parameters

Ring-1T is InclusionAI's (Ant Group's) open-weight trillion-scale thinking MoE from the Ling 2.0 family, trained from Ling-1T-base via long-CoT supervised fine-tuning, reasoning RLVR, and general RLHF. InclusionAI reports 1T total parameters and approximately 50B activated per token; the official checkpoint has 80 layers, GQA, 256 routed experts with eight selected per token plus one shared expert, QK normalization, half (partial) RoPE, and a 64K native context that can be extended to the reported 128K via YaRN. It targets extended natural-language reasoning and was evaluated across mathematics, coding, logic, healthcare, writing, and agent benchmarks.

Benchmark scores

BenchmarkScore
AA-LCR45.7
AA-Omniscience Index-47.9
AIME 202589.3
Artificial Analysis Intelligence Index16.2
Artificial Analysis Omniscience Accuracy19.8
Artificial Analysis Omniscience Hallucination Rate84.4
Artificial Analysis Openness Index38.9
Capability139.7
CritPt0.6
Global MMLU Lite79.8
GPQA Diamond77.4
Humanity's Last Exam (Text-Only)10.2
IFBench (Artificial Analysis)44.6
LiveCodeBench (Artificial Analysis)64.3
MMLU Pro80.6
SciCode (Artificial Analysis)36.7
SnakeBench Average Apples0.1
SnakeBench Best Apples1.0
SnakeBench Rating15.6
SnakeBench Total Apples1.0
SnakeBench Win Rate43.8
Terminal-Bench Hard6.8
τ²-bench (Telecom)26.3
Loading Atlas data…