Atlas

Models

← All models

K2 Horizon 375B A23B

LLM360 · 379.2B parameters

The Institute of Foundation Models' flagship K2 Horizon sparse checkpoint, published through the K2/LLM360 program. It has native 524,288-token context and request-level reasoning effort controls; the complete BF16 checkpoint contains 379,167,159,168 parameters. The card reports approximately 23B active parameters per token, recorded at that published precision.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo1298
AA-Briefcase v1.1 Elo1293
AA-LCR v1.180.0
AA-Omniscience Index-3.0
AA-Omniscience v1.1 Accuracy18.6
AA-Omniscience v1.1 Hallucination Rate29.0
AA-Omniscience v1.1 Index-5.0
Artificial Analysis Intelligence Index v4.330.8
Artificial Analysis Intelligence Index v4.3.230.5
Artificial Analysis Intelligence Index v5.017.8
Artificial Analysis Omniscience Accuracy18.2
Artificial Analysis Omniscience Hallucination Rate25.9
Artificial Analysis Progress Index v1294
AutomationBench-AA 1.0.637.2
Capability153.1
CritPt (avg@5)4.6
CritPt v2 (agentic, avg@5)29.7
FrontierCode 1.1 Main (Artificial Analysis)32.1
FrontierCode 2.0 (Artificial Analysis)11.8
GDP.pdf (Artificial Analysis, All-pass)7.4
GDP.pdf (Artificial Analysis, Mean Criteria)54.0
GDPval-AA v21400
GDPval-AA v2.11366
GPQA Diamond (Artificial Analysis avg@5)87.3
Humanity's Last Exam (Text-Only)32.0
Humanity's Last Exam Diamond10.2
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)12.2
Medical Long Context Reasoning (MLCR-AA)6.7
MLCR-AA Accuracy (Judged Responses)68.0
MLCR-AA Accuracy (October 2026 methodology)47.2
Loading Atlas data…