Atlas

Models

← All models

Kimi K3.1 Pro

Moonshot AI · 2026-10-24

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, released on October 24, 2026 (Beijing time) as a post-trained version of Kimi K3 with substantially larger reinforcement-learning compute across hundreds of downstream domains. Moonshot calls it its most capable 3T-class open-weight model, with a 1-million-token context window, and serves it at launch through Kimi.com, Kimi Work, Kimi Code and the Kimi API as kimi-k3.1-pro, alongside a separately named Kimi K3.1 Pro Highspeed endpoint (kimi-k3.1-pro-highspeed) said to run 2-3x faster. Launch results use max reasoning effort with temperature 1.0 and top-p 0.95, and the API documents low, high and max reasoning effort. Moonshot says it removed cyber-related data from mid-training and RL for the open-weight release and will publish the full weights by November 4, 2026; it has not yet released them, a license or an exact parameter count.

Benchmark scores

BenchmarkScore
AA-Briefcase v1.1 Elo1678
APEX-Agents 1.1 Pass@167.2
AutomationBench-AA 1.0.673.3
BabyVision with Python95.4
BrowseComp92.1
Capability165.9
Chartography90.2
CharXiv Reasoning89.0
CharXiv Reasoning (with tools)93.4
DeepSWE v1.175.5
ExploitGym - release-reported percentage, cohort unspecified20.7
FrontierFinance - release-reported version unspecified58.6
GDPval-AA v2.11846
GPQA Diamond94.2
Humanity's Last Exam70.0
Humanity's Last Exam Diamond39.4
Kimi Code Bench v277.1
MMMU Pro84.1
MMMU-Pro with Python90.0
OfficeQA Pro67.4
PerceptionBench68.3
ProgramBench - almost solved rollout percentage31.5
SWE-Marathon - release-reported version unspecified51.7
Terminal-Bench 4.065.7
Terminal-Bench-Science 0.138.1
VoxelBench Text Rating2265
ZeroBench (Main questions, pass@5, with tools)63.0
ZeroBench (Main questions, pass@5)39.0
ZeroBench Main Mean pass@1 (five samples)29.8
Loading Atlas data…