Moonshot AI logo

Kimi K3.1 Pro

Multimodal
Moonshot AI

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, a post-trained version of Kimi K3 built with substantially larger post-training compute and large-scale joint reinforcement learning across hundreds of domains. Moonshot describes it as its most capable 3T-class open-weight model, with a 1,048,576-token context window, text/image/video input, tool calling, structured output, and always-on reasoning (low, high, or max; Moonshot API default max). max_completion_tokens defaults to 131,072 and can be raised to 1,048,576. It targets long-horizon coding, 3D creation and CAD, financial research, and other agentic work. Pricing is a one-third cut from K3 ($2 / $10 vs $3 / $15 per MTok). A Highspeed variant (kimi-k3.1-pro-highspeed) offers 2-3x the speed with the same capabilities at $4 / $20 (cached input $0.20). Full weights are scheduled for release by Nov 4, 2026.

Key Specifications

Parameters
-
Context
1.0M
Release Date
October 23, 2026
Average Score
65.9%

Timeline

Key dates in the model's history
Announcement
October 23, 2026
Last Update
October 31, 2026
Today
November 20, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Fine-tuned from
kimi-k3
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$2.00
Output (per 1M tokens)
$10.00
Max Input Tokens
1.0M
Max Output Tokens
1.0M
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis
GPQA
GPQA-Diamond; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
94.2%

Other Tests

Specialized benchmarks
AA-Briefcase v1.1
Elo score; cited from public leaderboard results reported in the Moonshot blog • Self-reported
55.9%
Agents' Last Exam
Official v1.1 release, all 122 tasks; Kimi Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
43.2%
APEX-Agents
Cited from Mercor APEX-Agents leaderboard • Self-reported
67.2%
AutomationBench-AA
Cited from public leaderboard results reported in the Moonshot blog • Self-reported
73.3%
BabyVision
With Python; average of 4 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
95.4%
BenchCAD
Without tools; single-shot with official BenchCAD prompts; 1,000-case subset; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
72.8%
BenchCAD (with Python tool)
With tools; Claude Code harness; 1,000-case subset; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
92.7%
BioMysteryBench
Official v11 release; Kimi Code harness; Kimi K3 judge; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
80.6%
BrowseComp
Context compaction at 200K tokens; max 500 steps; temperature 1.0, top-p 1.0 • Self-reported
92.1%
Chartography
With Python; average of 5 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
90.2%
CharXiv-R
Reasoning questions; average of 4 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
89.0%
DeepSWE
DeepSWE v1.1; Claude Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
75.5%
ExploitBench
Official harness; 300-turn budget; pass@5 max capability score; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
74.7%
ExploitGym
KimiCode harness; 6-hour wall-clock budget per task; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
20.7%
GDP.pdf
Strict pass rate; page images plus Docling text; Gemini-3.5-Flash judge; average of 5 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
34.6%
GDPval-AA 2.1
Elo score; cited from public leaderboard results reported in the Moonshot blog • Self-reported
61.5%
Harvey's Legal Agent Benchmark
All-pass rate on public 1,251-task set; Stirrup harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
23.7%
Humanity's Last Exam
HLE-Full (2,500 questions) without tools; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
51.1%
Humanity's Last Exam (with tools)
HLE-Full (2,500 questions) with tools; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
70.0%
Kimi Code Bench v2
Internal benchmark; Claude Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
77.1%
MMMU-Pro
Official protocol; average of 4 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
84.1%
MMMU-Pro (with tools)
With Python; official protocol; average of 4 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
90.0%
OfficeQA Pro
Claude Code harness; PDFs rendered as images only; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
67.4%
PerceptionBench
Internal benchmark; average of 4 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
68.3%
PostTrainBench
PostTrainBench v1.2; Claude Code harness; H20 GPUs; DeepSeek-V4.1-Flash-1017 judge; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
36.5%
SWE-Marathon
SWE-Marathon v1.1; Claude Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
51.7%
Terminal-Bench 4.0
Claude Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
65.7%
Terminal-Bench-Science 0.1
Claude Code harness; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
38.1%
ZEROBench
Main split; pass@5; 5 runs; Max reasoning effort; temperature 1.0, top-p 0.95 • Self-reported
39.0%

License & Metadata

License
unknown
Announcement Date
October 23, 2026
Last Updated
October 31, 2026

Compare Kimi K3.1 Pro

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.