Kimi

There are two things to consider: whether you need an agent that delivers a finished file, and whether you are comfortable with an English interface.

Who it is for. For people who work with long documents and presentations: you upload a hundred-page report, get a summary, ask for slides by morning. The free tier is enough to tell whether this is your tool.

Who else. For people who want a strong model for free: at launch, Kimi K3 ranked third in the Artificial Analysis index after Claude Fable 5 and GPT-5.6 Sol (Jul 17, 2026), while ranking first among open models in its class (AA card, Sep 2, 2026), and it is available in chat at no cost.

Who it is not for. The swarm of agents is available only on paid plans.

A good fit: section 3 will take you from sign-in to your first task. Choosing between it and DeepSeek: the choice is in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

Three levels: regular chat, an agent that creates files, and an agent swarm for large tasks.

Chat. A question, web search, or review of an uploaded file or image. Two response modes, fast Instant and thoughtful High, plus a context window of up to one million tokens, meaning several thick books in one conversation. One million is available from the Allegro plan; the free plan has a shorter context window.

Agent. Describe a task in words, and Kimi creates the result as a file: a presentation, a spreadsheet with formulas, a document, a report of ten thousand words, or a working website. The vendor lists more than twenty tools for the agent and connects financial-data plugins such as Wind and S&P Global.

Swarm. For a large task, the main agent breaks it down and assigns it to up to 300 copies of itself, which work in parallel and take up to 4,000 steps. The vendor promises results 4.5 times faster than one agent. The swarm is available from the Moderato plan, where the limit is more modest than in the ads: from 2 subagents on Moderato to 8 on Allegro; the free plan gets the regular agent.

Kimi Work and Kimi Code. Desktop Kimi Work edits files directly on your computer: in the official video, it receives an Excel financial model and rewrites it using an annual report. Kimi Code lives in terminals and code editors, from VS Code to Cursor, and its source code is open.

What was actually done. In the official footage: McKinsey-style slides on fusion energy, a voxel 3D game from a single sentence, a copy of a site made from a video recording. We have not put Kimi through a task of our own, the numbers come from official demos and other people's tests.

The product has two addresses.

Open kimi.ai: this is the international version. The address kimi.com is made for China, where sign-in works only through WeChat or a phone with the +86 code, and the site itself sends foreigners to kimi.ai.

Step 1. Ask your first question without registering: the chat responds anyway, and sign-in is only needed for history and synchronization.

Step 2. To save history, sign in through a Google account. The +7 code is not listed among countries (checked Sep 2, 2026), and there is no email sign-in button in the window.

Step 3. Switch to the agent: Slides, Docs, Sheets, Websites, and Deep Research buttons are directly under the input field. Start the first task with the result name: "create a 10-slide presentation based on this PDF."

Apps. iPhone: App Store. Android: Google Play, official package com.moonshot.kimichat. Desktop apps for macOS and Windows are available from the website; full Kimi Work runs on Macs with Apple silicon.

Steps verified Sep 2, 2026 using live kimi.ai and kimi.com sign-in windows, the vendor help center, and the T-J guide

Two signals line up: the model is strong and the audience is large.

Artificial Analysis index. At launch, Kimi K3 scored 57 points and ranked third after Claude Fable 5 and GPT-5.6 Sol, level with Opus 4.8 and GPT-5.5 (Jul 17, 2026), while its index task cost $0.94 versus $1.80 for Opus 4.8. By Sep 2, 2026, the live Artificial Analysis model page shows 60 points for K3 max mode and first place out of 111 among large models with open weights. The first open model in its class, but not the overall leader.

Vendor claims. Vendor claims. On the model's GitHub card (captured Sep 2, 2026): GPQA Diamond 93.5, DeepSWE 67.5, BrowseComp 91.2. There is no independent reproduction; practitioners on Habr explicitly complain about the lack of third-party benchmarks, so treat these figures as self-reported.

Audience. Audience. The latest official figure is old: founder Yang Zhilin reported 36 million monthly active users in December 2024, while third-party 2026 snapshots show a decline and fourth-to-eighth place in the Chinese market behind Doubao and DeepSeek. Similarweb counted 60.1 million visits to kimi.com over the three months ending in July 2026, rank 852 worldwide, with 42 % of traffic from China.

Kimi's place among the assistants in our catalog is shown under the card title, and the full tool ranking sits on the shared page.

You do not have to pay for chat and the agent; payment is for parallelism, the swarm, and file storage.

Free entry. The Adagio plan provides K3 chat and an agent: one task at a time, two projects, two schedules, 500 MB. It does not include Swarm or Kimi Code.

Paid tiers. From $19 per month for Moderato to $199 for Vivace, from $15 with annual billing. You cannot buy now: as of Sep 2, 2026, the pricing page has a notice about new plans, separating Kimi and Kimi Code into individual subscriptions, and every paid-plan button leads to a waitlist.

For developers. The API costs $3 per million input tokens and $15 per million output tokens; cached context repeats cost $0.30. There is no free tier, and the minimum top-up is $1. The three subscriptions are separate: chat, Kimi Code, and API do not pay for one another.

Three questions: where your data is stored, what the model refuses to discuss, and how much to trust its answers.

Data in China. Moonshot AI is registered in Beijing, and kimi.com's footer lists Chinese ICP registration numbers. The full privacy policy was not reviewed for this assessment, so the practical rule is simple: do not upload contracts, personal data, or trade secrets to the chat, as with any foreign service.

Censorship. The only independent run found, ellamind's test across 168 topics sensitive to the PRC, gave Kimi K2.5 two failures, while DeepSeek V3.2 failed 136. This is one source with an unverified methodology; Western press, however, reports built-in topic controls in all Chinese models.

Answers need fact-checking. Reports and slides can look polished even when the numbers are wrong, so double-check the figures.

What else to know. In summer 2026, the head of the White House science office accused Moonshot of training K3 on Claude Fable 5 answers, which is called distillation. No evidence was presented; researchers object that two weeks passed between the releases of Fable and K3. The company removes old models from the API without a transition period: k2.5 has returned an error since 31.08.2026.

The choice is not “who is smarter,” but “what output do you need.”

Different workflows. DeepSeek is stronger at coding and math and remains a chat: question, answer. Kimi handles long documents and ready-made files: slides, tables, websites, and reports produced by a swarm of agents, while desktop Kimi Work edits files directly on the computer.

Different entry barrier. DeepSeek signs you up by email, Kimi needs a Google account.

Who else is nearby. ChatGPT remains the second reference point: a Plus subscription at $20 a month.

Over three years, a long-context chat became an agentic work environment, while models have moved toward open access with each generation.