# Kimi K3.1 Pro: Moonshot's open multimodal model, what it measures and when to use it

> Kimi K3.1 Pro from Moonshot AI: open weights, 1M context, image input, the benchmarks its lab published and its license, as of November 15, 2026.

*Tags: ai, models, moonshot, kimi, open-weights*

---


> [!info] Quick Definition
> **Kimi K3.1 Pro** is Moonshot AI's flagship model: an open model in the 3-trillion-parameter class (2.8T total and 104B active according to its card), multimodal (it accepts text and images) and with a 1M-token context window, aimed at coding, agents, 3D and finance. Moonshot calls it its most capable open-weight model of that class to date.


Data is as of November 15, 2026 and comes from Moonshot pages. The price does not appear in the announcement or the models documentation, so no figure is given here.

## What is Kimi K3.1 Pro and when did it come out?

Moonshot presents it as its most capable open model. The announcement shows no publication date; it does say the full weights would be released by November 4, 2026 at the latest, and they are now available on Hugging Face.

| Fact | Value |
|---|---|
| API ID | `kimi-k3.1-pro` (variant: `kimi-k3.1-pro-highspeed`) |
| Weights | `moonshotai/Kimi-K3.1-Pro` (Hugging Face) |
| Parameters | 2.8T total, 104B active (mixture of experts) |
| Context | 1M tokens (1,048,576) |
| Inputs | Text and image |
| Price | Not in the announcement or the models documentation: check Moonshot's pricing page |
| Open or closed | Open weights, its own license |

The API offers compatibility with the OpenAI and Anthropic formats, which makes it easy to try from tools you already use.

## What is it good for?

The announcement highlights three areas: coding (from quick fixes to full projects), 3D creation and spatial reasoning, and end-to-end financial workflows. Because it accepts images, you can try it on screenshots, charts and PDFs; Moonshot does not describe a specific chart-reading feature, only benchmark results.

Moonshot also reports two internal tests that matter for your world: DECK-Arena, which according to the lab evaluates presentation generation on real-world queries, and Design Bench, which covers graphic design with direct prompts or with reference images. These tests are designed and run by Moonshot; the score it publishes is a signal, not an independent verdict. To understand this kind of model, read [[multimodal-models|multimodal models]] and [[reasoning-models|reasoning models]].

## Which benchmarks did it publish and how do you read them?

Moonshot publishes all of these scores and, unless stated otherwise, they use reasoning effort max, temperature 1.0 and top-p 0.95 (BrowseComp used top-p 1.0). For the general method, read [[ai-benchmarks|how to read AI benchmarks]].

| Benchmark | Score | Who measured it and how |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Moonshot, with the Claude Code harness |
| DeepSWE v1.1 | 75.5 | Moonshot, with the Claude Code harness |
| BrowseComp | 92.1 | Moonshot (context compaction at 200K tokens, 500-step maximum) |
| OfficeQA Pro v2 | 64.4 | Moonshot (PDFs as images with no machine-readable text; tasks whose answer was found online scored zero) |
| GPQA-Diamond | 94.2 | Moonshot |
| HLE-Full with tools | 70.0 | Moonshot |
| Kimi Hallu Bench Hard (internal), with tools | 8.3 | Moonshot; lower is better |
| Kimi Hallu Bench (internal), without tools | 5.4 | Moonshot; lower is better; a separate row, not the Hard version |

How to read them:

- **Look at the harness.** Moonshot ran Terminal-Bench 4.0 with the Claude Code harness and takes other models' scores from blogs or public leaderboards. Changing the harness changes the number. Alibaba, for example, reports Terminal Bench 2.1, which is another version.
- **Internal tests.** Kimi Hallu Bench Hard measures the rate of hallucinated claims, assessed by fact-checking agents, on an internal set of challenging prompts drawn from work tasks; the lab itself says it is a stress test and the rates do not represent the overall error rate in daily use.
- **Jumps between versions.** The post compares with the previous K3 (for example, the hallucination rate fell from 23.4% to 8.3%). That is an improvement declared by the lab.

## How much does it cost and how much context does it have?

Context is 1M tokens according to Moonshot's documentation. The price per million tokens does not appear in the announcement or in the platform's models page; both point to the pricing page. Before budgeting, open it and note the date. To estimate a cost you need [[tokens|tokens]] and [[cost-per-token|cost per token]], and for context, [[context-window|context window]].

The `kimi-k3.1-pro-highspeed` variant promises 2 to 3 times the speed with the same intelligence; its price does not appear either. In Kimi Enterprise, the announcement says an administrator must enable the model, because it is off by default.

## When should you pick it and when not?

Pick it when:

- You want an open model that accepts images and works for code or office documents.
- You need an API compatible with OpenAI or Anthropic formats to try it quickly.
- You want to download the weights, as long as you comply with the license (see [[open-vs-closed-models|open vs. closed models]]).

Avoid it when:

- You need a firm price today: neither the announcement nor the models documentation gives one.
- You run a service that offers the model to third parties with high revenue: the license asks for an agreement with Moonshot above 20 million dollars in 12 months.
- You want the cheapest model for simple tasks: a 2.8T-parameter model is a lot for that. See [[small-vs-large-models|small vs. large models]].

## Example

Illustrative example: a designer has a screenshot of a metrics dashboard and wants an HTML prototype. She uploads the screenshot to `kimi-k3.1-pro` and asks: "Reproduce this dashboard in HTML and CSS with the same hierarchy; list separately what you could not infer from the image."

What to try tomorrow: repeat the same screenshot with `reasoning_effort` set to `low` and to `max`, and compare time, tokens used and fidelity. Check the result by eye and with browser tools; a model's visual reading is an estimate.

## Common mistakes

- Taking for granted a price you saw in a review: confirm it on Moonshot's official page.
- Comparing the 65.7% on Terminal-Bench 4.0 with a score from another version of the benchmark.
- Treating the score of an internal test (DECK-Arena, Design Bench) as a standard.
- Leaving `max` reasoning on for trivial tasks.
- Forgetting that, to pass from one turn to the next with tools, the card asks you to send back the full assistant message with its reasoning and tool calls.

## Sources

- [Moonshot AI: Kimi K3.1 Pro announcement](https://www.kimi.com/en/blog/kimi-k3-1-pro)
- [Hugging Face: Kimi-K3.1-Pro](https://huggingface.co/moonshotai/Kimi-K3.1-Pro)
- [Hugging Face: Kimi-K3.1-Pro license](https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE)
- [Kimi platform: models documentation](https://platform.kimi.ai/docs/models)


---

Source: https://www.fernandoux.com/en/wiki/ai/kimi-k3-1-pro/
