Skip to content
trending

AI Tools

Kimi AI: What It Is, Kimi K3, Pricing & My Test (2026)

Kimi from Moonshot AI in October 2026: K3 model with 2.8 trillion parameters, plans from $0 to $199, API, Polish language support, and our test.

By 9 min read
Kimi AI: What It Is, Kimi K3, Pricing & My Test (2026)

👁 120 reads

Kimi is an AI chat and family of language models from Chinese company Moonshot AI in Beijing. The latest model, Kimi K3 from July 16, 2026, has 2.8 trillion parameters and open weights - the largest open model ever released. The chat at kimi.com has a free plan, paid plans start at $19 per month, and the API costs $3 per million input tokens.

Status as of October 3, 2026

  • Latest model: Kimi K3 (July 16, 2026) - 2.8 trillion parameters, 104 billion active, 1,048,576 token context, text and image, reasoning always enabled.
  • Chat plans: Adagio free, Plus $19, Pro $39, Max $99, Ultra $199 per month. In late September, Moonshot replaced the old musical names (Moderato, Allegretto, Allegro, Vivace) with these.
  • API: $3.00 per million input tokens, $15.00 per million output, $0.30 for cache reads.
  • App: Kimi 3.1.3 in App Store (September 23), 4.3 rating with 46.2k reviews on Android. Interface without Polish language.
  • Our test: Kimi K3 scored 89.2 out of 100 on six tasks in Polish - third place behind two DeepSeeks, ahead of Qwen and GLM.

What Is Kimi AI

Kimi is three things under one name. First, a chat in the browser (kimi.com) and in iOS and Android apps, similar to ChatGPT: conversation, web search, reading files and images, an agent that creates websites and presentations on its own. Second, the Kimi family of language models, the latest of which have open weights on Hugging Face. Third, Kimi Code, an agent for programming in the terminal and editor.

Behind it stands Moonshot AI, a startup founded in March 2023 in Beijing by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, with Alibaba and Tencent among investors. The chat launched publicly on November 16, 2023, and today operates in two versions: Chinese and international (kimi.com). We wrote about valuation and IPO plans in the articles who is Moonshot AI and Moonshot AI wants $50 billion valuation.

How to Get Started Step by Step

  1. Go to kimi.com or download the Kimi app (publisher: Beijing Moonshot Technology) from the App Store or Google Play. Watch out for fakes with similar names.
  2. Create an account. The free Adagio plan doesn’t require purchasing a subscription.
  3. You can write in Polish - the interface is in English (or one of 15 other languages), but the model responds in the language of the question.
  4. For longer tasks, enable agent mode: Kimi searches the web, writes code, creates a website or slides on its own. On the free plan, the agent executes one task at a time.
  5. If you program, install Kimi Code CLI: in PowerShell irm https://code.kimi.com/kimi-code/install.ps1 | iex, and on macOS and Linux curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash. Then in the project directory kimi and /login. Details in the Kimi Code guide.
  6. If you prefer API, create an account at platform.kimi.ai and choose the kimi-k3 model. The API is compatible with OpenAI and Anthropic formats. The same model is also on OpenRouter as moonshotai/kimi-k3.

Kimi Models

Model Release Parameters Context Purpose
Kimi K1.5 January 20, 2025 not disclosed - first model with reasoning
Kimi K2 July 2025 1 trillion, 32 billion active 128k agent, code, open weights
Kimi K2.5 January 2026 1 trillion 256k multimodal
Kimi K2.6 April 2026 - 262,144 general, text, image, video
Kimi K2.7 Code June 12, 2026 - 262,144 programming
Kimi K3 July 16, 2026 2.8 trillion, 104 billion active 1,048,576 flagship: code, agents, document work

K3 is a mixture-of-experts model: out of 896 experts, 16 work on each token. Moonshot built it on its own Kimi Delta Attention mechanism, which saves memory with long context, and trained it directly in 4-bit MXFP4 precision. Reasoning cannot be disabled - the reasoning_effort parameter accepts values low, high, and max, with max as default. For multi-turn conversation via API, you must send back the model’s entire previous response, including the reasoning trace.

K3 weights are on Hugging Face under the Kimi K3 license. This is essentially an MIT license with two conditions: a company with a “model as a service” offering and revenue above $20 million in 12 months must sign a separate agreement with Moonshot, and a product with over 100 million monthly active users or $20 million monthly revenue must display the name “Kimi K3”. For internal use, there are no conditions. Running it at home is out of the question: this requires server equipment, not a graphics card. What you can run locally, we described in the local AI guide.

What the Benchmarks Say

In the producer’s table, K3 in the max variant has, among others, 93.5% in GPQA Diamond, 88.3% in Terminal-Bench 2.1, 91.2% in BrowseComp, and 84.8% in OSWorld-Verified, which is on par with Claude Fable 5 and GPT-5.6 Sol. These are Moonshot’s numbers, not independent measurements. Artificial Analysis on launch day gave it 57 points in its index (then fourth place globally and first among open models). After changing to index version 4.3, K3 has 44 points in the max variant, with GLM-5.3 and Qwen3.8 Max at 45 points alongside it. The same page reports 35 tokens per second and an increase in hallucination rate in the knowledge test from 39% (K2.6) to 51%. We collect current facts about the model itself, with results from the Arena, in a separate hub Kimi K3, and a comparison with Anthropic’s model is on the page Kimi K3 vs Claude Fable 5.

The launch was accompanied by considerable turmoil: tech stock sell-off, suspension of new subscriptions due to lack of computing power, and allegations that K3 was built on Claude’s responses.

How Much Does Kimi Cost

Kimi has two price lists: a subscription for chat, agent, and Kimi Code, and separate API paid per token. We don’t include App Store prices because they vary by region and aren’t direct exchange-rate conversions.

Plan Website (USD monthly) Annual (USD) What You Get
Adagio 0 0 chat, 1 agent task at a time, 2 scheduled tasks
Plus 19 180 K3, more agent credits, 2 tasks at a time, Kimi Code
Pro 39 372 more credits, K3 with 1M context in Kimi Code
Max 99 948 chats up to 1M tokens
Ultra 199 1,908 highest limits, 4 agent tasks at a time

USD prices according to the kimi.com membership page and its readings from September 28-29, 2026. Details on what to choose and how not to overpay are laid out in the article Kimi pricing.

Kimi K3 API costs $3.00 per million input tokens and $15.00 per million output (before tax). Older models are cheaper: K2.6 and K2.7 Code at $0.95 / $4.00. On OpenRouter, K3 is offered by over 20 providers, the cheapest starting at $0.99 for input, but in 4-bit quantization. Hence our automatic notes about price changes, e.g., Kimi K3: price up 366% - they track the cheapest provider, not Moonshot’s price list.

Kimi in Polish

The app interface doesn’t have Polish - the language list in the App Store is English, Chinese, and 14 others. The model itself writes in Polish well. In our test (details below), the email to a building manager from Kimi K3 had the cleanest language of five models: no calques, no typos, correct declension. Worse with rules: the text was 3 words shorter than requested, and it added the proposed date below the signature.

It corrected 14 errors flawlessly, knew Polish facts (19 out of 20 answers), but when asked about the current president, it answered “Andrzej Duda” with the caveat that it might not know the 2025 election results. This is a clue that K3’s knowledge about Poland ends before August 2025. In translation from English, idioms came out naturally (we tested it in Polish), with one meaning error.

Our Test: Kimi K3 vs GLM, Qwen and DeepSeek

On October 3, 2026, we gave Kimi K3 and four other Chinese models the same six tasks in Polish through OpenRouter, with reasoning level low and provider pinned to Moonshot AI. Full description, responses, and rubric are in the article Kimi K3, GLM-5.3, Qwen3.8 Max and DeepSeek in Polish: test. Kimi’s results:

Task Kimi K3 Best in Test
Writing email (0-10) 8.5 Kimi K3 - 8.5
Correcting 14 errors 14/14 4 models - 14/14
Code: VAT ID and declension (29 tests) 29/29 all - 29/29
Logic and calendar (4 questions) 3/4 DeepSeek V4.1 Flash and V4 Pro - 4/4
Knowledge about Poland (20 questions) 19/20 GLM, Qwen, DeepSeek V4.1 Flash - 20/20
Translation with idioms (0-10) 8 DeepSeek V4.1 Flash and V4 Pro - 9.5
Score (0-100) 89.2 DeepSeek V4.1 Flash and V4 Pro - 93.3
Time for 8 queries 144 s GLM-5.3 - 77 s
Cost for 8 queries $0.0719 GLM-5.3 - $0.0227

Two observations. At low level, K3 barely thinks on simple tasks (4-12 reasoning tokens for email, correction, code, and translation), so it responds in 8-15 seconds. It thought longer only on the logic puzzle: 1,403 tokens and 55 seconds. The only counting error was November 2026: it gave 19 working days instead of 20.

Kimi vs Competition

Kimi K3 GLM-5.3 Qwen3.8 Max DeepSeek V4.1 Flash Claude Opus 5.5
Company Moonshot AI Z.ai (Zhipu) Alibaba DeepSeek Anthropic
Weights open, Kimi K3 license open, GLM-5.3 license closed open, MIT closed
API: input / output (USD per 1M) 3.00 / 15.00 1.40 / 4.40 2.00 / 6.00 0.15 / 0.60 4.00 / 20.00
Artificial Analysis index (version 4.3) 44 45 45 - 58
Our test in Polish (0-100) 89.2 80.1 82.5 93.3 not tested in this series
Free chat yes, kimi.com yes, chat.z.ai (GLM-5.3-Flash) yes, chat.qwen.ai yes, chat.deepseek.com yes, with limit

API prices from producers, Claude according to OpenRouter, status as of October 3, 2026. Claude Opus 5.5 index according to reading from September 27. Guides to the others: GLM, Qwen, DeepSeek, Claude Opus 5.5.

Who Is Kimi For

  • For programmers who want a powerful model cheaper than Claude. The Plus plan for $19 gives K3 in Kimi Code, and the same model can be connected to other tools.
  • For companies that want to run a large model on their own. Open weights and license without restrictions for internal use, but a server with multiple GPUs is needed.
  • For writing in Polish via API - K3 writes cleanly and quickly, though DeepSeek V4.1 Flash performed better in our test at a fraction of the cost.
  • Not for sensitive data without legal department approval. Chat and API are servers of a company from China. In the US, there’s ongoing discussion about banning Chinese models, and researchers have shown that Kimi’s safeguards can be bypassed.

Frequently Asked Questions

Kimi AI - What Is It?

Kimi is an AI chat and family of language models from Moonshot AI in Beijing. It works in the browser at kimi.com and in mobile apps. The latest model is Kimi K3 from July 2026, with open weights.

Is Kimi Free?

Yes, the Adagio plan is free: chat and one agent task at a time. Moonshot doesn’t clearly state whether the free plan uses the full K3. Kimi Code and higher limits require a plan starting at $19 per month.

Does Kimi Work in Polish?

The model understands and writes in Polish well - in our test it had the cleanest language in an email among five Chinese models. The app interface doesn’t have a Polish version.

How Much Does Kimi K3 Cost via API?

$3.00 per million input tokens and $15.00 per million output at Moonshot AI, $0.30 for cache reads. On OpenRouter, some providers charge less but use 4-bit quantization.

Kimi or DeepSeek - Which Is Better?

In our test in Polish, DeepSeek V4.1 Flash scored 93.3 points, Kimi K3 89.2 points, and DeepSeek was three times cheaper. Kimi has better language and stronger results in agent tasks according to the producer.

Is Kimi Safe?

Chat and API send data to Moonshot AI servers. For company data, choose a version run on your own or at an EU provider. Open weights themselves don’t send anything.

Is Kimi K3 Open Source?

Weights are open on Hugging Face under the Kimi K3 license, similar to MIT. Restrictions apply only to large companies: with revenue above $20 million from “model as a service” offerings, an agreement with Moonshot is needed.

How to Download the Kimi App?

From the App Store or Google Play - search for the “Kimi” app from Beijing Moonshot Technology (on Android: Moonshot AI). The app is free, you buy paid plans inside.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

Kinetyra

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editor and publisher · Quantslant

I edit and publish Quantslant. AI tools help with research, drafts and translation; choosing topics, running the tests, checking facts and the decision to publish are mine, and so is the responsibility for what appears here.

read nextLMArena: How to Read the AI Model Ranking and Where It FailsguidesLMArena: How to Read the AI Model Ranking and Where It FailsLMArena (now Arena) ranks AI models by human votes, not test scores.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages6

Quantslant in numbers 1,423posts in archive22news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›