Kimi K3.1 Pro benchmarks, pricing and open weights

Kimi K3.1 Pro is Kimi K3 with more post-training

Kimi K3.1 Pro is a large language model that the Chinese AI company Moonshot AI announced on October 24, 2026, as the first model in its new K3.1 series. It went live the same day in four places:

• Kimi.ai, the chat service
• Kimi Work, the desktop app
• Kimi Code, the coding tool
• the Kimi API for developers

It is not a new architecture: Kimi K3.1 Pro is the July release of Kimi K3, post-trained with much more compute. Post-training is the stage after pretraining that turns a model into one people can actually use, and it is where skills such as coding or financial analysis get sharpened. Moonshot says it rebuilt its reinforcement learning (RL) setup, including the RL framework, long-context inference and training efficiency, so that it could train across hundreds of domains at once.

Moonshot calls K3.1 Pro its most capable "3T-class" open-weight model. An open-weight model is one whose trained weights are published, so anyone can download it and run it on their own servers. The K3 architecture and license are covered in Kimi K3 benchmarks, pricing and open weights.

The base Kimi K3 has 2.8T parameters, 104B active per token

Kimi K3 is a mixture-of-experts (MoE) model, which means it holds many specialized sub-networks and switches on only a few of them for each input. So although the base K3 has 2.8 trillion parameters in total, only 104 billion of them work on any single token. A token is the smallest unit of text a model processes, roughly a short word or part of one.

Its context window, the amount of text the model can take in at once, is 1,048,576 tokens, or about one million. Because K3.1 Pro is a post-trained K3, it keeps the same one-million-token window.

A Highspeed version runs at 2-3× the speed

On the same day Moonshot also released Kimi K3.1 Pro Highspeed, which answers two to three times faster than K3.1 Pro. On the API the two models are called kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and API customers can turn on Zero Data Retention so that their data is not stored.

In Kimi Enterprise, K3.1 Pro starts switched off, and an administrator has to enable it for the organization. Kimi Work needs desktop app version 3.1.0 or later, and in Kimi Code you pick the model with the /model command.

Benchmarks put it just behind Claude Opus 5.5 on coding

A benchmark is a shared set of test problems used to score and compare AI models. On those tests, Kimi K3.1 Pro lands just behind Claude Opus 5.5 on coding and ahead of GPT-6 Astra. Every number below comes from Moonshot's own blog, and unless noted, K3.1 Pro was run with reasoning effort set to max.

Kimi K3.1 Pro against other vendors' models

Bars show the score in percent (higher is better), scaled to the highest bar in each chart. Source: Moonshot AI blog.

Terminal-Bench 4.0

65.7Kimi K3.1 Pro

66.4Claude Opus 5.5

58.2GPT-6 Astra

0204060

BenchCAD (with tools)

92.7Kimi K3.1 Pro

96.2Claude Opus 5.5

95.9GPT-6 Astra

0255075

It beats GPT-6 Astra with 65.7% on Terminal-Bench 4.0

Terminal-Bench 4.0 tests how well an AI agent handles real coding tasks in a terminal, the text window where you type commands. K3.1 Pro scored 65.7%, ahead of GPT-6 Astra at 58.2% and just short of Claude Opus 5.5 at 66.4%.

Moonshot ran K3.1 Pro through Claude Code as the harness, the program that runs the model through each test. On Kimi Code Bench 3.0, its internal test of long, multi-step coding work, K3.1 Pro rose from K3's 55.95% to 61.2%.

It does not lead everywhere, though. On BenchCAD, which asks a model to rebuild 3D objects from several rendered views by writing CAD code, K3.1 Pro scored 92.7% with tools, behind Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Moonshot also says the model builds games in Unity, Unreal Engine and Godot, and writes scripts that drive Blender and CAD software.

Claim-level hallucinations drop from 23.4% to 8.3%

A hallucination is when an AI states something false as if it were true. On Kimi Hallu Bench Hard, Moonshot's internal test of factual accuracy, the share of individual claims that were hallucinated fell from 23.4% for K3 to 8.3%. The share of answers with at least one factual error fell to 42.4% of K3's rate.

The test is built from questions chosen because models tend to get them wrong. Moonshot notes that the rates are not the error rates you would see in everyday use.

Finance work improved as well. On OfficeQA Pro v2, which asks questions about real financial documents, K3.1 Pro scored 64.4, a result Moonshot calls state of the art. On its internal finance tests, financial modeling rose 7.2 points over K3 and financial research 11.6 points, and in a simulated business it scored about 4.2 times what K3 did.

API prices are a third below K3's, and the weights are out

An API is the interface that lets software call another service's functions and data. On the Kimi API, Kimi K3.1 Pro costs a third less than K3 for both input and output. The weights are also out on Hugging Face, so you can run the model on your own servers instead of going through Kimi's apps or the API.

Kimi API price per million tokens (US dollars)

Bars are scaled to the highest price in each chart. Prices from the Kimi API pricing page on October 29, 2026.

Input

2.00K3.1 Pro

4.00K3.1 Pro Highspeed

3.00K3

01234

Output

10.00K3.1 Pro

20.00K3.1 Pro Highspeed

15.00K3

05101520

Input costs $2 and output $10 per million tokens

Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens. K3 costs $3 and $15, so the same workload on K3.1 Pro comes to two thirds of the K3 bill. Cached input, the part of a prompt that repeats an earlier request, drops to $0.20, against $0.30 on K3.

Highspeed costs $4 for input and $20 for output, exactly twice the K3.1 Pro rate. API use is pay-as-you-go and is billed separately from a Kimi membership or Kimi Code.

The released weights were trained without cyber-related data

Moonshot had promised to release the full K3.1 Pro weights by November 4, 2026. The Hugging Face repository was created on November 2, and as of November 6 anyone can download the weights.

Moonshot removed cyber-related data from both mid-training and reinforcement learning before releasing them. It says these targeted controls lower the risk of misuse while keeping the model useful for general work.

A vulnerability is a security weakness in software or a system. On xdev-bench-hard, Moonshot's internal test of writing exploits for known vulnerabilities, a version trained without the data removal solved 90.5% of tasks, while the released open-weight K3.1 Pro solved 81.0%.

Moonshot says this cyber capability still trails frontier proprietary models that have been public for three months or more. It also says it will not release open weights that push the frontier of cyber capability forward.

Switch the API to K3.1 Pro now, read the license to self-host

If you already use Kimi K3 through the API, setting the model name to kimi-k3.1-pro is enough to try a cheaper K3.1 Pro. K3.1 Pro is built on K3 and keeps the same one-million-token context window. On top of that, its price is a third lower and its scores on coding and factual accuracy are higher.

Teams still on the older Kimi K2 models have until the end of 2026, at the latest, to move. Kimi's API model list says the K2 series stopped accepting new accounts on October 31, 2026, while accounts that already use it can carry on for now.

That grace period is short. Moonshot plans to switch off every K2 model by the end of December, and one of them, kimi-k2.5, was already shut down on August 31. The list names kimi-k3.1-pro as the model to move to.

If you want to download the weights and run them on your own servers, read the Kimi K3.1 Pro License on Hugging Face first. Only the main K3.1 Pro weights were released; the Highspeed version has none.

The released weights were trained without cyber-related data, and they score lower on the exploit-writing test than a version trained with that data. For work that needs exploit knowledge, such as security testing, factor that gap into the choice of model.