kimi-k3 vs kimi-k2
Side-by-side comparison of kimi-k3 and kimi-k2: benchmarks, pricing, context window and capabilities. Both are accessible through Requesty's unified API. kimi-k3 outperforms kimi-k2 on 3 of 3 shared benchmarks.
kimi-k3
Input / 1M
$3.00
Output / 1M
$15.00
Context
1.0M
Model ID
moonshot/kimi-k3
kimi-k2
Input / 1M
$0.60
Output / 1M
$2.50
Context
262K
Model ID
vertex/kimi-k2
Benchmark comparison
Intelligence Indexreasoning
kimi-k350.2%
kimi-k225.9%
Coding Indexcoding
kimi-k376.2%
kimi-k2N/A
Math Indexmath
kimi-k3N/A
kimi-k294.7%
GPQA Diamondreasoning
kimi-k393.5%
kimi-k283.8%
AIME 2025math
kimi-k3N/A
kimi-k294.7%
LiveCodeBenchcoding
kimi-k3N/A
kimi-k285.3%
Terminal-Bench Hardagentic
kimi-k3N/A
kimi-k231.1%
τ²-Benchagentic
kimi-k3N/A
kimi-k293.0%
SciCodecoding
kimi-k359.5%
kimi-k2N/A
MMLU Proknowledge
kimi-k3N/A
kimi-k284.8%
Humanity's Last Examreasoning
kimi-k346.9%
kimi-k223.8%
Scores sourced from official model cards, Artificial Analysis, and public leaderboards. Benchmarks measure specific skills and don't capture every aspect of model quality.
Pricing & specifications
| kimi-k3 | kimi-k2 | |
|---|---|---|
| Input price / 1M | $3.00 | $0.60 |
| Output price / 1M | $15.00 | $2.50 |
| Context window | 1.0M tokens | 262K tokens |
| Max output | 1.0M tokens | 262K tokens |
| Vision input | Yes | N/A |
| Tool calling | Yes | Yes |
| Reasoning | Yes | Yes |
| Prompt caching | Yes | Yes |
| Computer use | N/A | N/A |
| Provider | Moonshot AI | Google LLC (Vertex AI) |
Questions people ask
Is kimi-k3 better than kimi-k2?
kimi-k3 outperforms kimi-k2 on 3 of 3 shared benchmarks. See the benchmark comparison above for specifics: kimi-k3 and kimi-k2 have different strengths across reasoning, coding, math and multimodal tasks.
Which is cheaper, kimi-k3 or kimi-k2?
kimi-k2 is cheaper. kimi-k3 costs $3.00/$15.00 per 1M input/output tokens, while kimi-k2 costs $0.60/$2.50.
Can I use kimi-k3 and kimi-k2 through the same API?
Yes. Requesty provides a single OpenAI-compatible API that routes to both. Change just the "model" parameter to switch between "moonshot/kimi-k3" and "vertex/kimi-k2", no other code changes needed.
What are the context windows?
kimi-k3 supports up to 1.0M tokens of context. kimi-k2 supports up to 262K tokens. Longer context means you can feed larger documents or codebases in a single prompt, though quality often degrades past 128K for most models.
Switch between kimi-k3 and kimi-k2 with one line of code
Requesty provides a single OpenAI-compatible API for 600+ models. Change the model parameter, not your code.