Atlas

Models

← All models

KAT-Coder-Pro V1

Kuaishou

KAT-Coder-Pro V1 is Kuaishou's proprietary agentic coding model for programming tasks and workflows. StreamLake documents parallel multi-tool invocation, code comprehension, and logical reasoning, along with function calling, a 256K-token context window, and up to 32K output tokens.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo602
AA-Briefcase v1.1 Elo602
AA-LCR74.0
AA-LCR v1.138.7
AA-Omniscience Index-37.7
AIME 2025 (Artificial Analysis avg@10)94.7
Artificial Analysis Omniscience Accuracy17.7
Artificial Analysis Omniscience Hallucination Rate67.4
Artificial Analysis Progress Index v1240
AutomationBench-AA9.6
Capability140.9
CritPt (avg@5)0.0
GDPval-AA v2845
GDPval-AA v2 (normalized)20.1
GDPval-AA v2.1739
Global MMLU Lite81.7
GPQA Diamond (Artificial Analysis avg@5)76.4
Humanity's Last Exam (Text-Only)33.6
IFBench (Artificial Analysis avg@5)68.4
LiveCodeBench (Artificial Analysis)74.7
MMLU Pro81.3
SciCode (Artificial Analysis avg@3)36.6
SnakeBench Average Apples6.0
SnakeBench Best Apples16.0
SnakeBench Rating18.0
SnakeBench Total Apples60.0
SnakeBench Win Rate60.0
Terminal-Bench Hard9.1
τ²-bench (Telecom)88.6
Loading Atlas data…