Atlas

Models

← All models

Ling 2.6 1T

InclusionAI · 2026-04-29 · 1T parameters

Ling 2.6 1T is InclusionAI's open-weight instant/instruct MoE flagship, released on April 29, 2026; the official FP8 checkpoint contains 1,025,414,425,344 parameters excluding quantization-scale tensors, while InclusionAI describes its activated size only approximately as 63B. It upgrades the Ling 2.0 base through architectural migration and about 9.6T tokens of additional pre-training, combining Lightning Attention and MLA in a 7:1 ratio with a native 262,144-token limit. Post-training emphasizes fast, token-efficient answers and agentic coding and tool-use workflows, making it the instant-response counterpart to the deeper-reasoning Ring 2.6 1T.

Benchmark scores

BenchmarkScore
AA-LCR34.7
AA-LCR v1.141.7
AA-Omniscience Index-50.6
ALE-Bench1108
Artificial Analysis Omniscience Accuracy21.9
Artificial Analysis Omniscience Hallucination Rate92.8
Artificial Analysis Openness Index38.9
Artificial Analysis Progress Index v1238
BuseyBench SVG2.3
Capability138.2
CritPt (avg@5)0.3
Extended NYT Connections 940 (Quadratic v1)4.1
Extended NYT Connections 940-Puzzle Set10.8
Extended NYT Connections Newest 10013.0
Extended NYT Connections Newest 100 (Quadratic v1)5.5
GPQA Diamond (Artificial Analysis avg@5)75.2
Humanity's Last Exam (Text-Only)8.7
IFBench (Artificial Analysis avg@5)56.9
ObviousBench68.8
SciCode (Artificial Analysis avg@3)37.0
SnakeBench Average Apples7.5
SnakeBench Best Apples20.0
SnakeBench Rating24.9
SnakeBench Total Apples75.0
SnakeBench Win Rate70.0
Terminal-Bench Hard31.1
τ²-bench (Telecom)89.8
Loading Atlas data…