What Kimi is and who makes it
Kimi is a chat assistant built by Moonshot AI, a Beijing startup founded in 2023 and named after the Pink Floyd album 'The Dark Side of the Moon'. It arrived in late 2023 with one very specific pitch: while everyone else was arguing about reasoning benchmarks, Moonshot built an assistant that could swallow an enormous amount of text in a single conversation and still answer questions about page 400 as accurately as page 4.
That remains the reason to use it. Kimi is not trying to be the friendliest chatbot or the best image generator. It is a reading machine. You give it a 300-page annual report, a bundle of court judgments, a full API documentation set or six months of WhatsApp exports converted to text, and it holds the whole thing in working memory rather than chopping it into fragments and hoping the right fragment gets retrieved.
Moonshot has since released the Kimi K series of models, including open-weight releases that developers can download and run themselves. That matters for Indian teams who want the capability without sending data to a Chinese server, and we will come back to it.
Why a long context window actually matters
Most people never hit a context limit, so the feature sounds abstract. It stops being abstract the moment you work with documents.
When an assistant runs out of context, one of two things happens. Either it silently forgets the earlier part of the conversation, or the product quietly switches to retrieval, where a search layer picks a few paragraphs it thinks are relevant and shows only those to the model. Retrieval is fine for 'what is the notice period clause'. It falls apart for 'find every place in these three contracts where the indemnity language is inconsistent', because that question requires the model to have seen all three contracts at once.
Kimi is built for the second kind of question. Cross-document comparison, tracking a definition across a long statute, spotting a clause that contradicts something 200 pages earlier, summarising a thesis chapter by chapter without losing the argument thread. These are the tasks where you notice the difference.
Key features in depth
Whole-document reading. Upload PDFs, Word files, spreadsheets and plain text and ask questions across all of them together. It handles scanned-looking PDFs less reliably than clean digital ones, so run OCR first if your file is a photocopy of a 1998 order sheet.
Web browsing and citation. Kimi can search the live web and quote what it found. Its search quality is noticeably better on Chinese-language sources than on Indian ones, which is worth knowing before you rely on it for, say, a GST notification lookup.
Agentic modes. Newer Kimi releases include a research mode that plans a multi-step investigation, opens several sources and writes a longer report at the end rather than answering in one shot. It is slower and occasionally wanders, but on a broad question it produces a more complete first draft than a single-turn answer.
Coding. The K2 generation is genuinely competent at code, particularly at reading an unfamiliar repository and explaining what it does. It is not the first tool a professional developer would reach for over the established coding assistants, but for understanding legacy code it is useful precisely because it can hold many files at once.
Open weights. Moonshot has published open-weight versions of its K2 models. If you have the hardware, or you rent GPUs from an Indian provider, you can run the model inside your own boundary. Very few teams will do this, but for a firm with genuine confidentiality obligations it converts Kimi from unusable to usable.
Pricing for Indian users
The consumer chat product is free. There is no rupee-priced subscription, no student plan, no Indian payment flow to worry about, because there is nothing to pay for. You sign in and use it, subject to rate limits that tighten when the service is busy.
The cost appears only if you use the API. Moonshot prices its API per million tokens in US dollars, in the same broad range as other efficient models rather than the premium frontier tier. As a rough working figure, a developer processing a million tokens of input might expect a bill of a few tens of rupees rather than a few hundred, though the exact number moves with the model tier and with any caching discount. Convert at roughly Rs 85 to the dollar and treat every figure as indicative only, because Chinese AI API pricing has been cut repeatedly and any number printed today ages badly.
The practical point for an Indian buyer is this: for individual use, Kimi costs nothing, which makes the comparison against a Rs 1,700 to Rs 2,000 per month Western subscription unusually stark. What you pay instead is in caveats, not currency.
Kimi versus Claude and ChatGPT
| What you care about | Kimi | Claude | ChatGPT |
|---|---|---|---|
| Cost for one person | Free | Free tier, paid plan around Rs 1,700 per month indicative | Free tier, paid plan around Rs 1,700 per month indicative |
| Very long documents | Its core strength | Also very strong | Good, historically the weakest of the three on sheer length |
| Hindi and Indic languages | Workable Hindi, weak on other Indian languages | Better Hindi, reasonable Indic coverage | Broadest Indic coverage of the three |
| Writing quality in English | Competent, occasionally stilted | Generally the best | Very good, more generic |
| Ecosystem and integrations | Thin | Growing | Largest by far |
| Data handling comfort for Indian firms | The main concern | Enterprise terms available | Enterprise terms available |
| Images, voice, apps | Minimal | Minimal | Extensive |
The honest summary is that Kimi wins on price and on raw document capacity, and loses on everything that surrounds the model: polish, integrations, Indic language depth and the institutional comfort of using a US-headquartered vendor.
Real workflows where Kimi fits
A law student in Delhi reading judgments. Drop in a set of Supreme Court judgments and ask for the ratio in each, then ask which of them conflict. Kimi holds all of them at once, so the comparison is real rather than a stitched-together guess. Verify every citation before you put it in a submission.
A research scholar working through literature. Twenty papers, one conversation, and a request to map which authors disagree about the same variable. This is the single most convincing demonstration of a long context window.
A startup founder in Indore doing due diligence. Three years of a target company's filings uploaded together, with a question about how the revenue recognition language changed over time. Do not upload anything commercially sensitive; use it on public filings.
A content team auditing an old site. Export a few hundred posts as text, ask for duplicate topics and contradictory claims, and get a consolidation plan.
A CA firm reviewing contracts. Genuinely useful, and genuinely the place to stop and think about confidentiality before uploading a client's agreement. Use the open-weight route or a Western vendor with enterprise terms for anything privileged.
Limitations to know
Hindi and Indic support is mediocre. Kimi will answer in Hindi, but the phrasing is often flat and it does not handle Hinglish, Devanagari transliteration or code-mixed input with the ease that ChatGPT or Gemini do. Tamil, Bengali, Marathi and Telugu output is weaker still. If your work is in an Indian language, this is a real reason to pick something else.
India-specific knowledge is patchy. It knows the broad shape of Indian tax and company law but will get details of recent circulars, state-level rules and Indian financial products wrong. Never treat it as authority on an Indian compliance question.
The interface and support are China-first. Parts of the product, help pages and error messages surface Chinese text. Nothing is broken, but it is not a polished Indian user experience.
Data residency under Chinese law. This deserves plain language rather than either alarm or hand-waving. When you use the hosted service, your prompts and uploads sit on servers governed by Chinese law, which includes national security and intelligence legislation obliging domestic organisations to cooperate with state authorities. That is a different legal environment from the one covering a US or EU provider, and it is a legitimate reason many Indian enterprises, law firms, listed companies and anyone handling government or defence-adjacent work restrict Chinese AI services outright. Several governments have limited these tools on official devices. None of this says the model is malicious. It says the legal exposure is different, and that difference is the thing you should be assessing, not the marketing.
Availability can wobble. Access from India works, but sign-in flows and rate limits change without notice, and the product has periodically been harder to reach.
How to get started
- Go to the official Kimi site in a browser. Avoid third-party mirrors and lookalike apps, which are common.
- Sign in. Email or a Google account generally works from India; if a flow asks specifically for a mainland Chinese mobile number, you are on the domestic version rather than the international one.
- Start with a low-stakes document. Upload a public annual report or a set of published judgments, not a client file.
- Ask a cross-document question rather than a summary question. Summaries are where every tool looks good; comparisons are where Kimi separates itself.
- Check the answer against the source. Long-context models still misattribute, and a confident page reference can be wrong.
- Decide your confidentiality line before you go further. Write down what you will and will not upload, and stick to it.
- If you need it for real work at scale, look at the API or the open-weight models rather than the free chat window, so you get stable limits and, in the open-weight case, control over where data sits.