Atlas

Models

← All models

Mercury 2.5

Inception Labs · 2026-09-08

Inception Labs' September 8, 2026 diffusion language model for text reasoning, coding and agent workloads, with parallel tool calling and schema-aligned JSON output. It supports tunable reasoning and advertises a 260K-token context with up to 65,536 output tokens. The exact parameter count and an unrounded context limit are not published in the reviewed model documentation.

Benchmark scores

BenchmarkScore
AA-Briefcase v1.1 Elo377
AA-LCR v1.171.7
AA-Omniscience Index-39.5
Artificial Analysis Intelligence Index v4.3.212.3
Artificial Analysis Omniscience Accuracy22.6
Artificial Analysis Omniscience Hallucination Rate80.3
AutomationBench-AA 1.0.62.4
Capability141.5
CritPt (avg@5)0.0
GDP.pdf (Artificial Analysis, All-pass)3.6
GDP.pdf (Artificial Analysis, Mean Criteria)47.6
GDPval-AA v2.1462
Humanity's Last Exam (Text-Only)11.8
SciCode 1.0.1 (Artificial Analysis)38.5
Terminal-Bench 4.0 (Artificial Analysis)0.0
Loading Atlas data…