# Kimi K2 Instruct

One trillion parameters, 32 billion active per token. Every model in the K2 lineage inherits this architecture.

- **Model ID:** `moonshotai/kimi-k2`
- **Type:** chat
- **Providers:** novita
- **Context window:** 131,072
- **Maximum output tokens:** 131,072
- **Pricing:** $0.57/1M input tokens, $2.3/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/kimi-k2

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'moonshotai/kimi-k2',
  prompt: 'Why is the sky blue?'
})
```

## About

Kimi K2 Instruct, released July 11, 2025, is a Mixture-of-Experts (MoE) language model from Moonshot AI.

**Sparse expert routing at 32B activation.** The full trillion parameters encode broad knowledge: programming languages, API conventions, domain facts, and tool-use patterns. At inference time, a routing mechanism selects roughly 32 billion parameters per token. Latency and compute cost stay comparable to a dense 32B model, while the knowledge base spans the entire trillion-parameter budget.

With 32B active parameters for reasoning depth and a full 1T parameter budget encoding broad tool-use and coding knowledge, K2 handles structured sequences of API calls, multi-step planning, and code synthesis.

Kimi K2 Instruct is available through AI Gateway at $0.57 per million input tokens and $2.3 per million output tokens.

AI Gateway routes K2 across Novita AI, giving you automatic failover across multiple providers.

## What to consider

K2 routes across Novita AI. Choose it when uptime and provider redundancy matter most.

## When to use

### Best For

- Agentic pipelines requiring structured sequences of API calls, data processing, and code synthesis
- Deployments where provider redundancy matters most
- Teams evaluating the K2 architecture for the first time who want the original baseline
- Workloads that benefit from trillion-parameter knowledge breadth at 32B-dense inference economics

### Consider Alternatives When

- You need deliberative chain-of-thought traces. Kimi K2 Thinking layers extended reasoning on top of this foundation
- Minimum latency is the goal. Kimi K2 Turbo is the speed-optimized variant
- You need the September 2025 checkpoint's expanded context and refined agentic training
- Visual or multimodal inputs are part of the pipeline. K2 processes text only

## Best for

- **Agentic pipelines:** Structured sequences of API calls, data processing, and code synthesis
- **Provider redundancy:** Deployments where failover across multiple providers matters most
- **K2 architecture baseline:** Teams evaluating the K2 architecture for the first time who want the original release
- **Broad knowledge at low cost:** Workloads that benefit from trillion-parameter knowledge breadth at 32B-dense inference economics

## Consider alternatives

- **Chain-of-thought traces:** Kimi K2 Thinking layers extended reasoning on top of this foundation
- **Minimum latency:** Kimi K2 Turbo is the speed-optimized variant
- **September 2025 checkpoint:** Use Kimi K2-0905 for expanded context and refined agentic training
- **Multimodal inputs:** K2 processes text only, so reach for a vision-capable model

## Frequently asked questions

### How does sparse routing translate to cost savings?

The full 1T parameters store broad knowledge, but only ~32B activate per token via the expert router. You pay compute proportional to a 32B dense model while drawing on knowledge encoded across the entire trillion-parameter budget.

### Why does base K2 list many providers on AI Gateway?

It was the first K2 variant adopted across providers, so routing across Novita AI reflects earlier integration. Later checkpoints and variants can have narrower provider sets.

### Is K2 text-only?

Yes. Kimi K2 Instruct accepts and produces text. Multimodal capabilities are not part of this release.

### What agentic patterns does K2 handle well?

Structured multi-step sequences: invoke an API, parse the response, branch on results, call a second API, and synthesize a final output. The function-calling interface in AI Gateway maps directly to these workflows.

### Can I bring my own provider credentials?

Yes. AI Gateway supports Bring Your Own Key for providers where you hold a direct account. BYOK requests are excluded from ZDR coverage.

## Links

- [Model page](https://vercel.com/ai-gateway/models/kimi-k2)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://novita.ai/models)
- [Provider pricing](https://novita.ai/pricing)