In early 2024, a Beijing startup called Moonshot AI went viral in China for one feature nobody else had nailed: a chatbot that could read an entire novel without losing the thread. That product was Kimi, and its ultra-long context window made it the country’s first genuinely useful long-document assistant. Two years on, Kimi has quietly become something much bigger — an open-weight, agent-driven powerhouse backed by Alibaba and Tencent, valued at roughly $18 billion, and used by more people than any other Chinese AI tool. It now matches frontier Western models on the benchmarks developers care about, while costing a fraction as much. This review covers the consumer Kimi app at kimi.com; the Moonshot API is reviewed separately.

The models in 2026

Kimi’s app is powered by Moonshot’s K2 family, and the pace of releases has been relentless. The current flagship, Kimi K2.6 (April 2026), is a one-trillion-parameter Mixture-of-Experts model that activates only about 32 billion parameters per request — frontier quality without frontier running costs. It ships with a 256K-token context window and is natively multimodal. On Moonshot’s own figures it ties OpenAI’s GPT-5.5 on the SWE-Bench Pro coding benchmark (around 58.6%) and leads on Humanity’s Last Exam with tools, which is a remarkable result for an openly downloadable model.

In June 2026 Moonshot followed up with Kimi K2.7 Code, a coding specialist that uses roughly 30% fewer reasoning tokens while scoring higher on Moonshot’s coding suites. The cheaper, slightly older K2.5 remains available for cost-sensitive use. One honest caveat: many of the headline numbers for the newest models are vendor-reported, and independent third-party benchmarks lag behind each release — so treat the freshest scores as directional rather than gospel. The defining fact across the range is that the K2 models are open-weight under a Modified MIT licence, downloadable from Hugging Face, which becomes important when we get to privacy.

Features

Chat, Deep Research & long documents

The core experience is fast and capable, and Kimi’s long-context heritage shows: it handles big documents, codebases and research dumps without losing track. Deep Research runs multi-step investigations and returns sourced reports, while OK Computer is its agentic mode for executing longer tasks. Bilingual output is a genuine strength — it drafts in English and Traditional Chinese in the same session, and the Chinese reads as written rather than translated.

Kimi Code, Slides & Websites

Kimi Code is an open-source CLI coding agent for your terminal, pitched directly against Claude Code and Aider, and it’s included from the entry paid tier. Kimi also generates Slides (presentations) and simple Websites from a prompt, putting it ahead of leaner rivals like DeepSeek on built-in tooling.

Agent Swarm & Kimi Claw

The headline feature is Agent Swarm, which coordinates up to 300 parallel sub-agents to chew through parallelisable work far faster than a single agent could. Kimi Claw goes further still — an “always-on” agent that can observe and act on your files, apps and activity continuously. It’s powerful, but as we’ll cover below, it’s also the feature that most concerns security researchers.

Pricing in 2026

Kimi names its tiers after musical tempos, and the entry paid plan sits right alongside ChatGPT Plus and Claude Pro — while the free tier is unusually generous. Note that exact USD figures vary by region and billing period, and mainland China is priced separately in RMB.

For developers, the price-to-performance ratio is hard to beat, and the open weights mean self-hosting is a real option. For everyday users, “free with no message cap” is a strong offer — just remember the free tier runs a standard model, credits are metered from a shared pool and unused ones can expire, so the genuinely impressive K2.6 sits behind a subscription.

Pros and cons

Where it falls short

The biggest concern is data privacy. Moonshot’s consumer privacy policy lists a Singapore-incorporated entity and never mentions China, but the company is Beijing-based, and China’s National Intelligence Law (Article 7) gives authorities broad power to compel data access from Chinese companies and their subsidiaries. The policy also allows user content — prompts, files and generated outputs — to be collected and potentially used to improve models. This isn’t just theoretical: in April 2026 Kimi reportedly disclosed one user’s private resume to an unrelated user during a routine task, an apparent data-isolation failure that the company initially attributed to “hallucination.” The sensible advice is the same as for any China-hosted tool — fine for general, non-sensitive work, but keep confidential, client or personal data off the hosted app. As with DeepSeek, the open weights are a genuine escape hatch: self-host a K2 model and the data-residency problem disappears.

Kimi Claw sharpens that concern. An always-on agent with continuous access to your files, apps and communications represents a far deeper level of data exposure than a normal chatbot, and security analysts have flagged it specifically. Then there’s censorship: in line with Chinese content rules, Kimi deflects on politically sensitive subjects and applies notably strict guardrails on sensitive history and NSFW content. For coding and everyday tasks this rarely surfaces, but if you value unfiltered inquiry it’s a real limit. Finally, the sheer pace of releases — three flagship-class models in roughly five months — makes it genuinely hard to know which model you’re using, and the support and ecosystem around Kimi remain thinner than the Western incumbents’.

Scorecard

How Kimi scores across what matters, averaging to 7.6 / 10:

Verdict

Kimi is one of the most impressive AI stories outside the US labs. It pioneered long context, then reinvented itself as an agentic, open-weight competitor that genuinely trades blows with the frontier on coding — and it does it cheaply, with a free tier far more generous than most. If you want a powerful, low-cost brain for technical work, long documents and multi-agent tasks, and especially if you can self-host the open weights, it’s an excellent choice that punches well above its price.

It scores 7.6 rather than higher because the trade-offs are real: the hosted app’s data sits under Chinese law, a documented 2026 incident showed user data leaking between accounts, sensitive topics are censored, and Kimi Claw’s always-on reach is a serious consideration. Use it freely for non-confidential work; keep sensitive data and unfiltered research on a tool you trust. For comparable capability with similar caveats, see our DeepSeek review and Qwen review; for a privacy-friendlier open-weight option there’s our Le Chat by Mistral review; and our Claude review, ChatGPT review and Gemini review cover the all-round leaders.

Frequently asked questions

Reviewed June 2026 by AINewsAndUpdates.com. Models, features and pricing change quickly — always confirm current details on the official site before relying on it.