---
description: "Kimi K3.1 Pro by Moonshot AI on Verda Inference: a model for long-running coding and agentic work, with a 1M-token context window, image input and reasoning."
revision_date: 04.11.2026
---

# Kimi K3.1 Pro

Kimi K3.1 Pro is Moonshot AI's flagship model for long-running coding, agentic work and reasoning. It reads text and images, and works with up to 1M tokens of context.

## Model details

| Specification | Value |
|---------------|-------|
| Model ID | `kimi-k3-1-pro` |
| Provider | Moonshot AI |
| Context window | 1,048,576 tokens |
| Input | Text, image |
| Output | Text |
| Features | Tool calling, JSON mode, reasoning |
| Reasoning effort | `none`, `low`, `high`, `max` (default `max`) |
| Quantization | MXFP4 weights, MXFP8 activations |
| Runs in | Germany |
| Weights | [moonshotai/Kimi-K3.1-Pro](https://huggingface.co/moonshotai/Kimi-K3.1-Pro) |
| License | [Kimi K3.1 Pro License](https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE) |

Prices and your rate limits are on the model's page in the Console. Image input works through the API; the Playground in the Console takes text only.

## Good for

According to Moonshot AI, Kimi K3.1 Pro is strongest at:

- **Long-horizon coding**: from quick fixes in real codebases to full-project builds
- **Agentic work**: planning, tool use and multi-step execution
- **Financial research**: analysis over long financial documents
- **3D creation and design**: 3D models and scenes from a text brief, sketch or reference image

For Moonshot's benchmark results, see the [model card](https://huggingface.co/moonshotai/Kimi-K3.1-Pro). Test it on your own tasks before you choose.

## Verified by Moonshot AI

Our deployment of Kimi K3.1 Pro passed the acceptance check of Moonshot's [Kimi Vendor Verifier](https://www.kimi.com/en/blog/kimi-vendor-verifier), which tests whether a provider serves the model correctly. Scores at `max` reasoning effort:

| Benchmark | What it checks | Score |
|-----------|----------------|-------|
| OCRBench | Reading text in images | 87.8 |
| MMMU-Pro Vision | Reasoning over images across subjects | 84.6 |
| Kimi Tool Call | Tool-call triggering and JSON schema accuracy | 100 |
| BEAM (1M) | Memory over conversations of about 1M tokens | 39.76 |
| DeepSWE | Coding-agent tasks in real repositories | 75.7 |

## Call the model

```bash
curl -X POST "https://inference.verda.com/v1/chat/completions" \
  -H "Authorization: Bearer <your key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3-1-pro",
    "messages": [{"role": "user", "content": "What is the best viewpoint in Helsinki?"}],
    "max_tokens": 8192
  }'
```

For Python, Node and coding agents, see the [Quickstart](https://docs.verda.com/inference/quickstart/) and [Coding agents](https://docs.verda.com/inference/coding-agents/).

## Reasoning

Kimi K3.1 Pro thinks before it answers. It returns its reasoning in `reasoning_content`, and reasoning counts as output tokens. To control how much it thinks, set `reasoning_effort` to `low`, `high` or `max`. The default is `max`. To answer without reasoning, set it to `none`.

At `max`, reasoning can use all of a small `max_tokens` and leave `content` empty, so leave room: a simple question can use 2,000 reasoning tokens.

In multi-turn conversations and tool calls, send back the complete assistant message, including `reasoning_content` and `tool_calls`, not only `content`.