Kimi K3
Moonshot AI · released Jul 16, 2026
Kimi K3 is a 2.8T parameter open-weight multimodal reasoning model from Moonshot AI. It is suited for complex coding, knowledge work, and long-horizon agentic workflows, and is particularly strong at...
Specification
- Context window
- 1.05M
- Max output
- 944K
- Knowledge cutoff
- Not stated
- Parameters
- 2779.9B
- Licence
- other
- Serving providers
- 20
- Moderated
- No
- Uptime
- 100.0%
Intelligence
43.6
94th percentile
Coding
76.2
Coding Index
Agentic
50.0
Agentic Index
Output speed
114 t/s
Median across providers
Latency
526ms
Time to first token
Cost per task
$1.22
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- GPQA Diamond93.5%
- AA-LCR (long context)88.7%
- SciCode59.5%
- Humanity's Last Exam46.9%
- τ²-bench (Telecom)46.0%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| GPQA Diamond | 93.5% |
| AA-LCR (long context) | 88.7% |
| SciCode | 59.5% |
| Humanity's Last Exam | 46.9% |
| τ²-bench (Telecom) | 46.0% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- 45.4
- GLM 5.3Z.ai44.8
- Grok 4.6xAI44.3
- Step 5 PreviewStepFun43.7
- Kimi K3Moonshot AI43.6
- Claude Haiku 5.5Anthropic43.4
- GPT-5.6 TerraOpenAI42.1
- GLM 5.3 FlashZ.ai41.8
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Qwen3.8 Max (0902) | 45.4 |
| GLM 5.3 | 44.8 |
| Grok 4.6 | 44.3 |
| Step 5 Preview | 43.7 |
| Kimi K3 | 43.6 |
| Claude Haiku 5.5 | 43.4 |
| GPT-5.6 Terra | 42.1 |
| GLM 5.3 Flash | 41.8 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.45
- Output / 1M tokens
- $15
- Cached input / 1M
- $0.28
- Blended 3:1
- $4.09
One task, estimated
$1.22
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
- Wafer
- Relace
- InferenceNet
- Morph
- Sail Research
- AkashML
- Makora
- Chutes
- Decart
- Phala
- DigitalOcean
- Parasail
- DeepInfra
- Amazon Bedrock
- Modal
- Together
- Fireworks
- BaseTen
- Moonshot AI
- Alibaba
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.
Arena
Head-to-head generation quality
Elo from pairwise judgements, broken out by the kind of thing the model was asked to build.
- Overall Elo
- 1,388
- Win rate
- 67.3%
- Strongest at
- 3D scenes
- Tournaments
- 3,364
Elo by category
Dot position on a 1,250–1,400 scale · Elo has no meaningful zero
- 1,388
- 1,364
- 1,351
- 1,346
- 1,336
- 1,299
- 1,315
- 1,264
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| 3D scenes | 1388 | 67.3% |
| Game development | 1364 | 60.9% |
| Data visualisation | 1351 | 63.2% |
| UI components | 1346 | 60.3% |
| Websites | 1336 | 59.2% |
| SVG | 1299 | 59.2% |
| Full-stack apps | 1315 | 61.8% |
| Mobile apps | 1264 | 57.0% |