Kimi K2.5
Moonshot AI · released Jan 27, 2026
Kimi K2.5 is Moonshot AI's native multimodal model, delivering state-of-the-art visual coding capability and a self-directed agent swarm paradigm. Built on Kimi K2 with continued pretraining over approximately 15T mixed...
Specification
- Context window
- 262K
- Max output
- 262K
- Knowledge cutoff
- Not stated
- Parameters
- 1058.6B
- Licence
- other
- Serving providers
- 10
- Moderated
- No
- Uptime
- 100.0%
Intelligence
36.0
83th percentile
Coding
46.8
Coding Index
Agentic
21.7
Agentic Index
Output speed
73 t/s
Median across providers
Latency
849ms
Time to first token
Cost per task
$0.26
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
- τ²-bench (Telecom)95.9%
- GPQA Diamond87.9%
- AA-LCR (long context)73.0%
- IFBench70.2%
- SciCode49.0%
- Terminal-Bench Hard34.8%
- Humanity's Last Exam30.7%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| τ²-bench (Telecom) | 95.9% |
| GPQA Diamond | 87.9% |
| AA-LCR (long context) | 73.0% |
| IFBench | 70.2% |
| SciCode | 49.0% |
| Terminal-Bench Hard | 34.8% |
| Humanity's Last Exam | 30.7% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Gemini 3.5 Flash LiteGoogle37.4
- Claude Sonnet 4.6Anthropic36.8
- Kimi K2.5Moonshot AI36.0
- Claude Opus 4.5Anthropic35.6
- GPT-5.1-CodexOpenAI35.6
- GLM 5V TurboZ.ai35.3
- GPT-5OpenAI35.3
- Qwen3.5-27BQwen34.6
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Gemini 3.5 Flash Lite | 37.4 |
| Claude Sonnet 4.6 | 36.8 |
| Kimi K2.5 | 36.0 |
| Claude Opus 4.5 | 35.6 |
| GPT-5.1-Codex | 35.6 |
| GLM 5V Turbo | 35.3 |
| GPT-5 | 35.3 |
| Qwen3.5-27B | 34.6 |
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
- $0.57
- Output / 1M tokens
- $2.85
- Cached input / 1M
- $0.095
- Blended 3:1
- $1.14
One task, estimated
$0.26
- 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
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
- 1,272
- Win rate
- 55.2%
- Strongest at
- Websites
- Tournaments
- 23,360
Elo by category
Dot position on a 1,150–1,275 scale · Elo has no meaningful zero
- 1,272
- 1,271
- 1,260
- 1,248
- 1,247
- 1,195
- 1,206
- 1,156
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| Websites | 1272 | 55.2% |
| UI components | 1271 | 53.6% |
| 3D scenes | 1260 | 53.1% |
| Game development | 1248 | 53.4% |
| Data visualisation | 1247 | 51.3% |
| SVG | 1195 | 48.4% |
| Mobile apps | 1206 | 54.2% |
| Full-stack apps | 1156 | 54.2% |