Atlas

Models

← All models

Qwen3.8 2.4T A95B

Alibaba

Qwen3.8 2.4T A95B is Alibaba's open-weight Qwen-Max-class mixture-of-experts checkpoint. This text-only artifact requires thinking but allows native low, medium, and xhigh effort, so Atlas classifies its variable reasoning as optional. It documents 262,144 native context tokens extendable to 1,010,000; the hosted Qwen3.8-Max service adds separate features.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo1439
AA-Briefcase v1.1 Elo1445
AA-LCR v1.180.3
AA-Omniscience Index4.3
Artificial Analysis Intelligence Index v4.340.0
Artificial Analysis Intelligence Index v4.3.239.9
Artificial Analysis Omniscience Accuracy31.3
Artificial Analysis Omniscience Hallucination Rate39.2
Artificial Analysis Openness Index27.8
AutomationBench-AA 1.0.657.2
BuseyBench SVG5.3
Capability160.1
Creative Writing v3 Elo1841
Creative Writing v3 Rubric Score16.7
CritPt (avg@5)20.0
EnterpriseOps-Gym-AA47.4
GDP.pdf (Artificial Analysis, All-pass)15.0
GDP.pdf (Artificial Analysis, Mean Criteria)64.6
GDPval-AA v21628
GDPval-AA v2.11597
GPQA Diamond (Artificial Analysis avg@5)93.5
Humanity's Last Exam (Text-Only)42.4
Medical Long Context Reasoning (MLCR-AA)0.0
MLCR-AA Accuracy (Judged Responses)59.0
MLCR-AA Completeness (Judged Responses)0.0
MLCR-AA Conciseness96.1
SciCode 1.0.1 (Artificial Analysis)54.1
SimpleBench62.5
Terminal-Bench 2.182.0
Terminal-Bench 4.0 (Artificial Analysis)11.1
Loading Atlas data…