Kimi K3
Kimi K3Category scores
Scores per capability area with the model's rank among ranked models in that category.
- 1Multimodal#192.6
- 2Knowledge#2383.9
- 3Agentic & tool use#1081.9
- 4Coding#2374.6
- 5Reasoning#1869.8
Benchmark results (52)
As reported in system cards, papers and public evaluations. Position compares this model with every other model reporting the same benchmark.
Multimodal
| Design Arena Website | 1336 | #3 of 91 |
| MMMU-Pro | 81.6 | #5 of 48 |
| CharXiv | 91.3 | #4 of 42 |
| ExtractBench (LlamaIndex)(lower is better) | 83.17 | #9 of 24 |
| MathVision | 94.3 | #2 of 20 |
| CharXiv w/o tools | 84.8 | #4 of 14 |
| OfficeQA Pro | 63.3 | #6 of 13 |
| MMMU-Pro w/ Python | 83.4 | #2 of 9 |
| ZeroBench | 23.0 | #6 of 8 |
| MathVision w/ Python | 97.8 | #1 of 5 |
| BabyVision w/ Python | 85.7 | #4 of 5 |
| ZeroBench w/ Python | 41.0 | #3 of 3 |
| PerceptionBench | 58.5 | #2 of 3 |
| WorldVQA ForceAnswer | 51.0 | #1 of 1 |
| OmniDocBench | 91.1 | #1 of 1 |
Agentic & tool use
| Terminal-Bench 2.1 | 88.3 | #7 of 68 |
| BrowseComp | 91.2 | #5 of 50 |
| MCP Atlas | 84.2 | #5 of 42 |
| MazeBench | 1 | #13 of 42 |
| JobBench | 52.9 | #13 of 34 |
| AutomationBench | 30.8 | #20 of 25 |
| Toolathlon-Verified | 73.2 | #12 of 22 |
| DeepSearchQA | 95.0 | #2 of 20 |
| ApprenticeBench | 18 | #9 of 18 |
| Hermes Index | 39.2 | #8 of 16 |
| Hermes Bench | 71.5 | #8 of 16 |
| APEX-Agents | 37.6 | #4 of 8 |
| SpreadsheetBench 2 | 34.8 | #1 of 1 |
| DECK-Bench | 73.5 | #1 of 1 |
Coding
| DeepSWE | 67.5 | #24 of 50 |
| Bug Hunt Bench | 21.0 | #18 of 23 |
| FrontierSWE v2 | 25.9 | #12 of 20 |
| ProgramBench | 77.8 | #6 of 15 |
| PostTrainBench v1.1 | 32.0 | #9 of 15 |
| VulcanBench v3 | 73.7 | #14 of 14 |
| OpenHarmony Bench | 57.3 | #6 of 12 |
| MLS-Bench Lite | 48.3 | #1 of 4 |
| FrontierSWE | 81.2 | #1 of 3 |
| PostTrain Bench | 36.6 | #2 of 3 |
| Kimi Code Bench v2 | 72.9 | #1 of 2 |
Knowledge
| GPQA | 93.5 | #7 of 117 |
| HLE | 56.0 | #10 of 69 |
| HLE w/o tools | 43.5 | #14 of 43 |
Reasoning
| ARC-AGI-2 | 60.4 | #27 of 41 |
| ARC-AGI-1 | 94.50 | #15 of 36 |
Instruction following
| Gray Swan IPI (15 attempts)(lower is better) | 52.7 | #12 of 12 |
Pricing by source
USD per 1M tokens.
| Source | Input | Output | Cached in |
|---|---|---|---|
| List price (BenchLM) | $3.00 | $15.00 | $0.30 |
| OpenRouter moonshotai/kimi-k3 | $0.24 | $11.94 | $0.21 |
Moonshot AI's official Kimi K3 pricing page lists $0.30 cache-hit input, $3.00 cache-miss input, and $15.00 output per million tokens for model ID kimi-k3, with an exact 1,048,576-token context window. OpenRouter independently lists route moonshotai/kimi-k3 with one Moonshot AI INT4 endpoint at the same rates and context length. Prices exclude applicable taxes. Moonshot schedules the full weight release for July 27, 2026, so source type remains pending until the files are public.
Arena ratings
Human-preference ratings from LMArena. Best-rated variant shown.
| Text(Max) | 1475 | #25 · 32.8K votes |
| WebDev (coding)(Max) | 1655 | #16 · 19.6K votes |
Speed
Measured by us through OpenRouter (median of trials).
Not measured yet.
Related models
Same family
- Kimi K2.7 Code49.4
- Kimi K2.658.4
- Kimi K2.553.0
Similar score
- Claude Opus 4.870.0 · $10.00
- Qwen3.8 Max69.7 · —
- GPT-5.6 Terra70.2 · $4.50
- Step 5 Preview69.6 · $1.43
- GPT-5.569.2 · $11.25