What it is
Kimi is the chat assistant from Moonshot AI. It runs on Kimi K3, a 2.8-trillion-parameter open-weight model launched in July 2026 with a context window of up to 1M tokens. Besides chat, Kimi offers agent tasks, Agent Swarm for parallel agents, Deep Research, document, spreadsheet and slide generation, website deployment, and the Kimi Code coding tool.
The consumer service is run by Beijing Moonshot AI, whose privacy policy says personal information is stored in mainland China. The developer API is run separately by Moonshot AI Pte. Ltd. in Singapore, on servers in Singapore.
Plans were renamed in 2026 from tempo names (Andante, Moderato, Allegretto, Allegro) to Go, Plus, Pro, Max and Ultra. Kimi's help centre lists them in yuan, from ¥49 to ¥1,399 a month, and K3 requires Plus or above.
Our review
Kimi's draw is the K3 model and the agent tooling around it. K3 has 2.8 trillion parameters and a 1M-token context, and on the API it costs $3 per million input tokens, $0.30 for cached input and $15 for output, flat across the whole window. Moonshot's launch post says its own API sees cache hit rates above 90 per cent on coding workloads, which brings the effective input price close to the cached rate for agent loops.
The plans are now legible, if only in yuan. Kimi's help centre lists Go at ¥49 a month, Plus at ¥99, Pro at ¥199, Max at ¥699 and Ultra at ¥1,399, with yearly billing cutting each by about a fifth. Go has no Kimi Code and no K3; Plus adds both plus Agent Swarm; the full 1M context and the high-speed coding model start at Pro. Every feature draws from one monthly pool that does not roll over, and Kimi Code adds a rolling five-hour throttle. USD prices on kimi.com were not retrievable when we checked, and subscribers on the old tempo-named plans keep their previous terms.
The consumer privacy policy is explicit. Personal information is stored inside the People's Republic of China, de-identified inputs and outputs are used to optimise models, and the only way to opt out is to contact customer service and verify your identity; there is no settings switch. Kimi Business, billed yearly with no discounts, promises that enterprise data is not used for training, but its FAQ names no ISO or SOC certification. The API is a different operation: a Singapore entity, no training on enterprise customer data by default, and zero data retention on request, excluding batch jobs and directly uploaded images and video.
Open weights change the calculation for engineering teams. K3 ships under the Kimi K3 licence with vLLM and SGLang recipes, but Moonshot recommends supernode configurations of 64 or more accelerators, so self-hosting is a cluster project rather than a server purchase. That leaves most teams choosing between Moonshot's Singapore API and a third-party host, and for a regulated workplace that choice matters far more than the chat app's slide generator.
Where it earns its place
A platform engineer at a 30-developer start-up
Run a coding agent on an open model without per-seat subscriptions
Point an OpenAI-compatible client at the K3 API: $3 per million input tokens, $0.30 on cache hits and $15 output. A $1 top-up activates K3, and rate limits rise with cumulative spend; risk-control throttles, once triggered, cannot be lifted.
An ML infrastructure team at a bank
Evaluate self-hosting K3 to keep code on premises
Weights are on Hugging Face under the Kimi K3 licence with vLLM and SGLang recipes. Moonshot suggests 64 or more accelerators, so budget a multi-node cluster and read the licence terms for large commercial deployments.
A China-based product manager at a consumer brand
Produce research briefs and slide decks in Chinese and English
Plus at ¥99 a month includes Deep Research, slides, Agent Swarm and K3, sharing one monthly pool that does not roll over. Ultra at ¥1,399 a month suits heavy parallel agent use.
Before you commit
- Setup
- Sign in on kimi.com or the iOS, Android and HarmonyOS apps; Kimi Code installs as a CLI with up to 5 API keys per member drawing on the same quota.
- Integrations
- OpenAI-compatible API with automatic context caching (5-minute or 1-hour TTL), batch API and web search endpoints; Kimi Code CLI works with membership or API keys.
- Data and privacy
- Consumer: stored in mainland China, de-identified chats used for training, opt-out by contacting support. API: Moonshot AI Pte. Ltd., Singapore servers, enterprise data not trained on by default, ZDR on request. No certifications published.
- Cost at scale
- An agent workload of 1 billion input tokens a month at a 90% cache hit rate and 100 million output tokens costs about $570 in input and $1,500 in output on the K3 API, before cache-write charges. Business seat prices are not published.
- Limits
- API tiers set concurrency, RPM, TPM and TPD by cumulative top-up, but the table did not render in the docs we read; exceeding them returns HTTP 429. Booster packs top up from ¥25, capped at ¥3,000 a day.
What's good
- Kimi K3 open weights, 2.8 trillion parameters, 1M-token context
- API at $3/$15 per 1M tokens with $0.30 cached input, flat across the window
- Agents, Deep Research, slides and Kimi Code under one quota
- Zero data retention available for API enterprise customers
What's not
- Consumer data stored in mainland China
- Consumer chats train models unless you ask support to opt out
- K3 needs Plus or above; full 1M context needs Pro
- Plans renamed in 2026; older tiers keep different rules
- No certifications named for Kimi Business
Best for
Developers and agent-heavy users outside regulated industries who want a frontier open-weight model with built-in research and coding tools.
Alternatives
Kimi: common questions
- Is Kimi worth it?
- Kimi wraps a frontier-class open model in capable agent tools, but the consumer app keeps your data in China and trains on it unless you ask support to stop. We rate it 3.7 out of 5.
- How much does Kimi cost?
- Free tier; Go ¥49, Plus ¥99, Pro ¥199, Max ¥699, Ultra ¥1,399 a month (¥39–¥1,119 billed yearly), per Kimi's help centre. Business annual only. API: K3 $3 input / $15 output per 1M tokens.
- What are the downsides of Kimi?
- Consumer data stored in mainland China; consumer chats train models unless you ask support to opt out; k3 needs Plus or above; full 1M context needs Pro; plans renamed in 2026; older tiers keep different rules; no certifications named for Kimi Business.
- Who is Kimi best for?
- Developers and agent-heavy users outside regulated industries who want a frontier open-weight model with built-in research and coding tools.
- How much does Kimi cost for a team or in production?
- An agent workload of 1 billion input tokens a month at a 90% cache hit rate and 100 million output tokens costs about $570 in input and $1,500 in output on the K3 API, before cache-write charges. Business seat prices are not published.
- Is Kimi safe for company data?
- Consumer: stored in mainland China, de-identified chats used for training, opt-out by contacting support. API: Moonshot AI Pte. Ltd., Singapore servers, enterprise data not trained on by default, ZDR on request. No certifications published.
- What are the best Kimi alternatives?
- The closest alternatives we have reviewed are DeepSeek, Qwen Chat and ChatGPT.
Answers checked Oct 23, 2026.