Atlas

Models

← All models

Ling 2.6 Flash

InclusionAI · 2026-04-28 · 107.5B parameters

Ling 2.6 Flash is InclusionAI's open-weight instant-response instruct MoE, released on April 28, 2026; InclusionAI describes its scale as about 104B total and 7.4B activated, while the official checkpoint metadata counts exactly 107,494,409,216 parameters. It retrofits Ling 2.0 with 32 layers of hybrid Lightning Attention and MLA in a 7:1 ratio and is optimized for token-efficient agentic tasks rather than Ring 2.6's distinct deeper-reasoning behavior. The checkpoint config is native to 131,072 positions, while the official SGLang instructions support 262,144 with a 2x YaRN override.

Benchmark scores

BenchmarkScore
AA-LCR25.0
AA-Omniscience Index-65.7
Artificial Analysis Agentic Index2.3
Artificial Analysis Coding Index25.3
Artificial Analysis Intelligence Index14.1
Artificial Analysis Omniscience Accuracy15.4
Artificial Analysis Omniscience Hallucination Rate95.8
Artificial Analysis Openness Index38.9
BuseyBench SVG1.5
Capability132.3
CritPt0.0
GDPval-AA v2550
GDPval-AA v2 (normalized)2.5
GPQA Diamond59.3
Humanity's Last Exam (Text-Only)6.2
IFBench (Artificial Analysis)57.4
SciCode (Artificial Analysis)27.1
SnakeBench Average Apples0.4
SnakeBench Best Apples1.0
SnakeBench Rating1.2
SnakeBench Total Apples4.0
SnakeBench Win Rate12.5
Terminal-Bench 2.124.3
Terminal-Bench Hard21.2
τ²-bench (Telecom)86.0
𝜏³-Banking2.9
Loading Atlas data…