Kimi

Kimi is an AI assistant and model family offered through a consumer site at kimi.com and a developer platform, the Kimi API Platform, with documentation at platform.kimi.ai. The consumer site navigation lists New chat, Plugins, Scheduled tasks, Slides, Deep Research, Docs and Projects, and its headline announces K3.1 Pro with upgraded coding and agentic capabilities.

The API platform describes K3.1 Pro as its flagship model for long-horizon coding and knowledge work, with a 1,048,576-token context window, and says it is an open-weight model of 3T class. Developers call it through a Chat Completions endpoint compatible with the OpenAI SDK. The model accepts image and video input, supports tool calling and structured JSON output, and always runs with thinking enabled. Older K3, K2.7 Code and K2.6 models remain in the model list.

• Long context. All K3 series models support a 1,048,576-token context window, and K2 series models 262,144 tokens. The vendor says pricing is flat, without tiers by context length.
• Reasoning effort control. Kimi K3.1 Pro has thinking always on. The reasoning_effort parameter accepts low, high or max, with max as the default, letting developers trade depth against latency and token use.
• Tool calling and structured output. The API supports tool calling with tool_choice, custom tools and dynamic tool loading, strict JSON Schema output, JSON mode and a partial mode that continues supplied text.
• Vision input. Images and videos can be passed as input. Public image URLs are not accepted; use base64 data or an ms:// file reference, and upload videos as files.
• Context caching. Repeated request prefixes are cached automatically with a default 5-minute TTL, with a 1-hour option for K3 models. Cached input is billed at a cache-hit price.
• Batch inference. K3.1 Pro and K3 support batch jobs for work that does not need real-time answers, billed at 60% of the standard price, per the vendor.
• Consumer workspace. The kimi.com site lists Deep Research, Slides, Docs, Projects, Plugins and scheduled tasks. The pages reviewed do not describe how each works.

• A developer team building a coding agent. Use K3.1 Pro with tool calling and the 1M-token window to feed an entire repository, and set reasoning_effort to low for quick edits and max for hard refactors. Batch inference suits overnight test generation.
• An analyst automating document review. Extract text from uploaded documents, pass it to the model with JSON Schema output, and cache the repeated instructions prefix. The vendor lists file extraction and storage as temporarily free.
• A small product team migrating from OpenAI. Because the endpoint follows Chat Completions and works with the OpenAI SDK (openai 1.0 or later), changing the base URL and model name is the main code change, followed by removing fixed sampling parameters.

For the API, the vendor's steps are: create an API key in the Kimi console, install the OpenAI Python SDK (openai>=1.0, Python 3.9+), create a client with base URL https://api.moonshot.ai/v1 and your key, then call chat.completions.create with model kimi-k3.1-pro. The key can be read from the MOONSHOT_API_KEY environment variable.

The model unlocks after a top-up of at least $1, and cumulative top-ups set your rate-limit tier. Fixed parameters such as temperature should be omitted, and multi-turn calls must return the complete assistant message. The consumer app has no setup steps described on the pages reviewed.

Kimi is run by Beijing-based Moonshot AI and sells memberships with a shared usage pool, top-up packs and a business edition with seat management.

The pricing page loads client-side, so we saw no prices, and the help centre we read is in Chinese. Its newest model, K3, is described as a 1M-token context model, which suits long documents more than quick questions.

The AI chatbots and assistants list places Kimi 7 of 14: DeepSeek sits directly above it and Grok directly below. Established names take the places after the first three, in ranked order, and we give no numeric score.

Kimi is based in China; 2 of the 14 products here share its region, Asia-Pacific.

In the catalogue: Membership tiers and business edition; prices not readable without JavaScript.

The API is pay-as-you-go with no subscription plan: input and output tokens are billed per million tokens (1 MTok = 1,000,000 tokens), and prices exclude taxes calculated at checkout. The per-model amounts are shown in tables that could not be read on the pages fetched, so no figures are quoted here. Printed terms: a minimum top-up of $1 unlocks K3.1 Pro, batch jobs cost 60% of the standard price, and file APIs are temporarily free. The consumer membership and Kimi Code are separate products, and their prices are not stated on the pages reviewed.

Basis: kimi.com, read for the Best AI Chatbots and Assistants in 2026 guide (reviewed 6 November 2026).

Kimi's help centre describes memberships with a shared usage pool, extra top-up packs and automatic renewal that can be cancelled.

The privacy agreement is issued by Beijing Moonshot Technology Co., Ltd.

Moonshot's own site says K3 has 2.8T parameters, is multimodal and has a 1M-token context.

Hosting region, certifications and data retention are not stated on the documentation pages reviewed. The pricing page says tax is calculated by jurisdiction at checkout. The platform lists a Kimi Business contact route for enterprise requests.

The vendor says K3.1 Pro is an open-weight model, but the API itself is a hosted service at api.moonshot.ai, so data handling terms should be read in the vendor's legal pages before sending sensitive content.

Integrations the vendor names (7):

• OpenAI SDK compatibility
• Chat Completions API
• Tool calling (Formula official tools)
• Token Calculation API
• Batch API
• File extraction API
• Kimi Code

4 products in this guide come closest to Kimi on tags, pricing and region.

• DeepSeek (#6 of 14): Chinese lab's chat app and web assistant, data processed in China. Watch out: the terms do not state a data storage location and are governed by PRC law.
• Google Gemini (#9 of 14): Google assistant tied to a Google account and storage tiers. (United States)
• Meta AI (#11 of 14): Meta assistant used inside its apps, with paid Meta One tiers. Suits: Consumers who already use Instagram, WhatsApp or Facebook every day.
• You.com (#12 of 14): Search company whose site now leads with web search APIs. Best for: Developers wanting real-time web search for their own AI tools. (United States)

Read from the vendor’s own pages on 5 November 2026.

• platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart.md: model capabilities, parameters, API steps, top-up rule
• platform.kimi.ai/docs/pricing/chat: billing logic, batch discount, K2 retirement
• platform.kimi.ai/docs/guide/product-plans.md: pay-as-you-go versus membership
• kimi.com/en: consumer product navigation and K3.1 Pro announcement

The same four questions are asked of every AI chatbots and assistants product: time to value, real cost, the way out and the intended buyer. The result for Kimi: #7, out of 14 ranked. Details: how we rank software.