Atlas

Models

← All models

Kimi K2 Thinking

Moonshot AI · 2025-11-06 · 1.1T parameters

Kimi K2 Thinking is Moonshot AI's open-weight thinking-agent checkpoint released in November 2025. It is a one-trillion-parameter MoE with 32B active parameters, a 256K-token context window, native step-by-step reasoning interleaved with tool calls, and an INT4 checkpoint produced with quantization-aware post-training.

Benchmark scores

BenchmarkScore
AA-LCR66.3
AA-Omniscience Index-20.5
Agent Red Teaming k=100 (CAIS)92.8
AIME 2024 (avg@8)85.8
AIME 2024-202585.4
AIME 2025 (avg@8)85.0
ALE-Bench1103
Apex (MathArena)0.0
Apex Shortlist47.3
APEX-Agents4.1
APEX-SWE11.2
ARC-AGI-25.0
Arena Code WebDev — Brand & Marketing1332
Arena Code WebDev — Consumer Product1318
Arena Code WebDev — Content Creation Tools1325
Arena Code WebDev — Data & Analytics1320
Arena Code WebDev — Gaming1310
Arena Code WebDev — HTML1326
Arena Code WebDev — React1316
Arena Code WebDev — Reference-Based Design1312
Arena Code WebDev — Simulations1302
Arena Text — Business, Management & Financial Ops (No Style Control)1397
Arena Text — Chinese (No Style Control)1453
Arena Text — Creative Writing (No Style Control)1375
Arena Text — English (No Style Control)1428
Arena Text — Entertainment, Sports & Media (No Style Control)1377
Arena Text — Exclude Ties (No Style Control)1400
Arena Text — Exclude Ties (Style Controlled)1423
Arena Text — Expert (No Style Control)1438
Arena Text — French (No Style Control)1425
Loading Atlas data…