Atlas

Models

← All models

Kimi K3

Moonshot AI · 2026-07-16

Kimi K3 is Moonshot AI's flagship reasoning model released on July 16, 2026, an approximately 2.8-trillion-parameter Stable LatentMoE transformer that activates 16 of 896 experts and supports a 1,048,576-token context with native image and video input. Its architecture combines Kimi Delta Attention, Attention Residuals, Gated MLA, and Sigmoid Tanh Unit, while Per-Head Muon and SFT-stage quantization-aware training with MXFP4 weights and MXFP8 activations support scaling; Moonshot reports roughly 2.5× better scaling efficiency than Kimi K2. K3 always reasons with low, high, or max effort (max at API launch), and Moonshot released the full model weights on July 27, 2026, under the Kimi K3 License.

Benchmark scores

BenchmarkScore
AA-LCR74.7
AA-Omniscience Index18.4
Agent Red Teaming k=100 (CAIS)76.6
Agents' Last Exam (ALE)28.3
Agents' Last Exam ALE-CLI Pass Rate27.6
Agents' Last Exam ALE-CLI Score52.8
Agents' Last Exam Full-Spectrum Pass Rate29.1
Agents' Last Exam Full-Spectrum Score49.9
Agents' Last Exam Last-Exam Pass Rate10.5
Agents' Last Exam Last-Exam Score20.6
Agents' Last Exam Near-Term Pass Rate40.3
Agents' Last Exam Near-Term Score71.6
Agents' Last Exam Overall Score51.6
Agents' Last Exam Unlicensed Full-Spectrum Pass Rate28.0
Agents' Last Exam Unlicensed Full-Spectrum Score48.1
Agents' Last Exam Unlicensed Last-Exam Pass Rate13.3
Agents' Last Exam Unlicensed Last-Exam Score23.1
Agents' Last Exam Unlicensed Near-Term Pass Rate40.3
Agents' Last Exam Unlicensed Near-Term Score71.6
Agents' Last Exam Unlicensed Overall Pass Rate29.8
Agents' Last Exam Unlicensed Overall Score53.0
AHK-Eval94.4
AHK-Eval Algorithms83.3
AHK-Eval Data Structures100.0
AHK-Eval Date and Time83.3
AHK-Eval Easy100.0
AHK-Eval Hard91.7
AHK-Eval Hidden Cases94.5
AHK-Eval Mid91.7
AHK-Eval Numbers100.0
Loading Atlas data…