Kimi K3
Moonshot · Released Jul 2026 · kimi-k3
$3.00
Input price / 1M tokens
#47 of 61, cheapest first
$15.00
Output price / 1M tokens
#46 of 61, cheapest first
1.05M
Context window
#11 of 71
34
Output tokens / second
#61 of 61
88.3%
Terminal-Bench 2.1
#4 of 17
67.5%
DeepSWE v1.1
#10 of 17
93.5%
GPQA Diamond
#5 of 27
43.5%
Humanity’s Last Exam (no tools)
#6 of 23
Summary
Kimi K3 is an open-weight model from Moonshot, released on Jul 2026.
At $3.00 input and $15.00 output per million tokens, it is #47 of 61 on input price, cheapest first.
Its 1.05M-token context window ranks #11 of 71.
Measured output speed is 34 tokens per second, #61 of 61.
Moonshot's flagship and the first open 3T-class model: 2.8T total, 104B active parameters, native vision, 1M context. Kimi K3 License allows commercial use, but Model-as-a-Service operators with over $20M revenue in 12 months need a separate agreement, and products over 100M MAU or $20M monthly revenue must display "Kimi K3".
Benchmarks · 6 reported
| Benchmark | Score | Bar | Rank | Source |
|---|---|---|---|---|
| Terminal-Bench 2.1max effort | 88.3% | #4 of 17 | Kimi-K3 model cardSelf-reported | |
| DeepSWE v1.1max effort | 67.5% | #10 of 17 | Kimi-K3 model cardSelf-reported | |
| GPQA Diamondmax effort | 93.5% | #5 of 27 | Kimi-K3 model cardSelf-reported | |
| Humanity’s Last Exam (no tools)full set, no tools, max effort | 43.5% | #6 of 23 | Kimi-K3 model cardSelf-reported | |
| Humanity’s Last Exam (with tools)full set, with tools, max effort | 56% | #11 of 16 | Kimi-K3 model cardSelf-reported | |
| OSWorld-Verifiedmax effort | 84.8% | #1 of 8 | Kimi-K3 model cardSelf-reported |
Self-reported means the lab ran the test itself. A rank appears only where several labs report the same version of a benchmark.
Similar models
Models with the most benchmarks in common with Kimi K3, and the closest scores.
Terminal-Bench 2.1
#4 of 17Terminal tasks, earlier set
- Kimi K388.3
- DeepSeek V4 Pro87.9
- DeepSeek V4.1 Flash90.6
- Qwen3.8-Max86.6
- GLM-5.281
DeepSWE v1.1
#10 of 17Long coding tasks in real repos
- Kimi K367.5
- DeepSeek V4 Pro62.7
- DeepSeek V4.1 Flash74.2
GPQA Diamond
#5 of 27Graduate-level science questions
- Kimi K393.5
- DeepSeek V4 Pro92.4
- DeepSeek V4.1 Flash90.9
- Qwen3.8-Max92.6
- GPT-5.593.6
- GLM-5.291.2
Humanity’s Last Exam (no tools)
#6 of 23Expert questions, no search or code
- Kimi K343.5
- DeepSeek V4 Pro42.7
- DeepSeek V4.1 Flash36.8
- Qwen3.8-Max43.6
- GPT-5.541.4
- GLM-5.240.5
Compare Kimi K3 side by side
Pre-filled with the two closest models. Swap any of them.
Sources
- price, cache write tiers, context: platform.moonshot.ai/docs/pricing/chat, checked 2 Oct 2026
- API launch month (July 2026): platform.kimi.ai/docs/platform-changelog, checked 2 Oct 2026
- max_completion_tokens up to 1,048,576, modalities: platform.moonshot.ai/docs/guide/kimi-k3-quickstart, checked 2 Oct 2026
- parameters, license, benchmarks: huggingface.co/moonshotai/Kimi-K3, checked 2 Oct 2026
- launch post, weights release by 27 Jul 2026: kimi.com/blog/kimi-k3, checked 2 Oct 2026
- output speed: artificialanalysis.ai/models/kimi-k3, checked 2 Oct 2026
Building on Kimi K3?
Get the architecture right before the bill arrives.
We size caching, routing and fallbacks for your workload on a free call, and tell you where a cheaper model is good enough.