# Kimi K3 Fast

Kimi K3 Fast runs Kimi K3 on a lower-latency serving path. Capabilities match the base model and the per-token rate is higher, so you pick it when response time shapes the experience.

- **Model ID:** `moonshotai/kimi-k3-fast`
- **Type:** chat
- **Providers:** fireworks, morph
- **Context window:** 1,000,000
- **Maximum output tokens:** 1,000,000
- **Pricing:** $4.5/1M input tokens, $22.5/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/kimi-k3-fast

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

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

const result = streamText({
  model: 'moonshotai/kimi-k3-fast',
  prompt: 'Why is the sky blue?'
})
```

## About

Kimi K3 Fast is the fast serving tier for Kimi K3, Moonshot AI's open-source model for long-horizon engineering work. Released July 27, 2026, Kimi K3 Fast trades a higher per-token cost for lower latency. The model's capabilities don't change.

Everything documented for the base model carries over. Kimi K3 Fast accepts text, image, and video input, supports a context window of 1M tokens, and keeps thinking mode always on. The strengths match too: long-horizon software engineering, knowledge work, deep reasoning, and tasks where code meets visual and spatial reasoning like frontend development, game development, and computer-aided design (CAD).

Two paths reach the fast tier. Set the `speed` option to `fast` while keeping the model ID on `moonshotai/kimi-k3`, and AI Gateway routes to the fast tier and falls back to standard speed when it isn't available. Or call `moonshotai/kimi-k3-fast` directly, which pins every request to the fast tier without that fallback.

Serving speed counts when someone is waiting. Interactive coding assistants stream output while a developer watches, and edit-run-fix loops repeat many times in a session, so per-turn latency compounds. Batch and offline jobs gain little. Throughput varies with load and context length, so read the live metrics on this page instead of planning around a fixed number.

AI Gateway routes Kimi K3 Fast across Fireworks, Morph, including US-based providers for teams with data residency and compliance requirements. To keep inference in US data centers, set `inferenceRegion` with scope `zone` and geo region `us`. Zero Data Retention is configured per request with `zeroDataRetention` or turned on for every request from AI Gateway dashboard settings, and support varies by provider.

Use the identifier `moonshotai/kimi-k3-fast` with the AI SDK or any supported interface like Chat Completions, Responses, or Messages. Kimi K3 Fast supports a context window of 1M tokens and completions up to 1M tokens per request, at $4.5 per million input tokens and $22.5 per million output tokens.

## What to consider

The fast tier lists above the base model, so compare $4.5 and $22.5 against Kimi K3 before you move high-volume traffic. Pinning `moonshotai/kimi-k3-fast` skips the fallback to standard speed that the `speed` option provides on the base model ID, so decide which behavior you want during a fast-tier shortage. Thinking mode stays always on, and those reasoning tokens still count toward the completion cap of 1M tokens. Measured throughput varies with load and context length, so use the live metrics on this page rather than a published figure.

## When to use

### Best For

- Interactive coding assistants where a developer watches output stream in
- Tight edit-run-fix loops where per-turn latency compounds across a session
- Long-horizon agents that need Kimi K3 quality under a response-time budget
- Teams already on Kimi K3 that want faster serving without changing model behavior

### Consider Alternatives When

- Cost per token dominates. Kimi K3 serves the same model at the standard rate
- The workload runs offline or in batch, where nobody waits on the stream
- You want automatic fallback to standard speed. Use the `speed` option on the base model ID
- Thinking mode must be optional. A non-thinking Kimi K2 variant generates answers directly

## Best for

- **Interactive Coding Assistants:** Streaming output to a developer who watches the session
- **Tight Iteration Loops:** Edit-run-fix cycles where per-turn latency compounds across a session
- **Latency-Budgeted Agents:** Long-horizon agents that need Kimi K3 quality under a response-time cap
- **Drop-In Speed Upgrade:** Teams on Kimi K3 that want faster serving with the same behavior

## Consider alternatives

- **Cost-Sensitive Traffic:** Kimi K3 serves the same model at the standard per-token rate
- **Batch and Offline Jobs:** Faster serving adds little when nobody waits on the output
- **Automatic Speed Fallback:** The `speed` option on the base model ID falls back to standard speed
- **Optional Thinking Mode:** A non-thinking Kimi K2 variant answers without a reasoning pass

## Frequently asked questions

### How is Kimi K3 Fast different from Kimi K3?

Serving speed and price. Kimi K3 Fast is the faster serving path for the same model, so the context window, native visual understanding, and always-on thinking all carry over. The per-token rate is higher; compare the rates and live metrics on each model's page.

### How do I request the fast tier?

Two ways. Set the `speed` option to `fast` while keeping the model ID on `moonshotai/kimi-k3`, which falls back to standard speed when the fast tier isn't available. Or call `moonshotai/kimi-k3-fast` directly to pin every request to the fast tier.

### How fast is Kimi K3 Fast?

Speed varies with load, context length, and routing, so a single number would mislead. Check the live throughput and latency metrics on this page for measured performance through AI Gateway.

### Does Kimi K3 Fast accept images and video?

Yes. Text, image, and video input all work, matching the base model. Visual understanding is native, so screenshots and screen recordings feed a coding loop without a separate vision model.

### Can I turn off thinking mode on Kimi K3 Fast?

No. Thinking mode is always on for the Kimi K3 models, so every response includes a deliberation pass. Budget output tokens for it, and reach for a non-thinking Kimi K2 variant when deliberation doesn't help.

### Can I keep Kimi K3 Fast inference in the United States?

Yes. AI Gateway serves the Kimi K3 models from US-based providers, and `inferenceRegion` restricts a request to US data centers. Regional inference lists above the standard rate, so check the pricing panel on this page first.

### How do I use Kimi K3 Fast on AI Gateway?

Use the identifier `moonshotai/kimi-k3-fast` with the AI SDK or any supported interface like Chat Completions, Responses, or Messages. AI Gateway routes across `fireworks`, `morph` and handles failover automatically.

### Does Kimi K3 Fast support zero data retention?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/kimi-k3-fast)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)