Kimi K2 Thinking: What Moonshot’s open model does, and what happened to its API

Plans, prices and features checked on the official pages on October 23, 2026.

Kimi K2 Thinking is an open-weight reasoning model from Moonshot AI, a Chinese AI company, released in November 2025. It is a mixture-of-experts model with 1 trillion total parameters that reasons step by step while calling tools. Its weights are still on Hugging Face, but according to Moonshot’s model list the hosted kimi-k2-thinking API model was retired in May 2026, so new API work should target the newer Kimi models.

This guide covers what the model does, how to use it today, what its successors are, and where pricing stands. Prices checked on October 23, 2026: Moonshot’s official pricing table did not display figures when read, so this guide gives no dollar amounts and points you to the live page instead.

What Kimi K2 Thinking is and who made it

Kimi is Moonshot AI’s model family and chatbot. K2 Thinking is the reasoning variant of the Kimi K2 line: instead of answering at once, it works through a problem, then calls tools such as search or code execution when needed, and continues reasoning with the results.

The Hugging Face model card, published by Moonshot, lists these specs:

• 1 trillion parameters in total, with 32 billion active per token.
• 384 experts, 8 selected per token, plus 1 shared expert.
• A 256K-token context window.
• Native INT4 quantization, which Moonshot says gives roughly 2x faster generation. The benchmark numbers it reports were measured at INT4.
• A modified MIT license covering both code and weights.

The model does not generate images or video. It is a text model that reasons and calls tools, so questions like “can Kimi K2 make pictures” have a plain answer: not on its own. The later K2.6 and K3 models accept image input, but that is understanding, not generation.

What it does well: Agentic reasoning and long tool chains

Moonshot’s pitch for K2 Thinking is agent work. The model card says it stays coherent across 200 to 300 consecutive tool calls, which matters for research tasks and multi-step coding where weaker models lose the thread.

The scores below are from Moonshot’s own model card, not independent tests:

Read that last row carefully: on this coding benchmark the card itself shows K2 Thinking behind both rivals. Vendor-reported numbers also depend on the test setup, so treat them as a rough guide. For how the first Kimi K2 compared with Qwen 3 and Claude Code on coding, see our comparison of China’s AI coding models.

Is the Kimi K2 Thinking API still available?

Not from Moonshot. The platform’s model list says the kimi-k2 series, including kimi-k2-thinking and kimi-k2-0905-preview, was discontinued on May 25, 2026 and is no longer maintained or supported. The changelog gives the same picture: K2 Thinking and a Turbo variant launched in November 2025 and were retired in May 2026.

The same pages list other retirements:

• kimi-k2.5: released in January 2026, closed to new users in July, and discontinued on August 31, 2026. The changelog says calls now return a 404 “model not found” error.
• moonshot-v1 models: discontinued on August 31, 2026.
• kimi-k2-thinking-turbo, kimi-k2-turbo-preview and kimi-k2-0711-preview: listed as deprecated.

Moonshot recommends kimi-k3 as the replacement for all of them. If an old tutorial tells you to call kimi-k2-thinking on Moonshot’s endpoint, expect it to fail.

The models that replaced it, and what the API costs

These are the active models on Moonshot’s platform as listed on its model page and quickstart guides:

On cost, Moonshot’s pricing page states that inference is pay-as-you-go with no subscription, billed per 1M tokens for input and output, with taxes added at checkout. For K3, cache writes are billed by time-to-live (5 minutes by default, or 1 hour), and cache hits are billed at the cached input price. The file extraction and storage APIs are described as temporarily free, though extracted text passed to a model counts as input.

The per-model rates themselves are on Moonshot’s pricing page. Third-party trackers quote different figures and some are out of date, so confirm there before budgeting. For a cross-vendor view, see our AI API pricing tracker.

How to get started with the API

1. Create an account on the Kimi open platform and add credit, since billing is pay-as-you-go.
2. Generate an API key in the console and store it as an environment variable, not in code.
3. Use the OpenAI SDK and set the base URL to https://api.moonshot.ai/v1, as the K2.6 quickstart describes. Moonshot also lists an Anthropic-compatible API.
4. Pick a model name: kimi-k3 for the flagship, kimi-k2.6 for a cheaper general model, or kimi-k2.7-code for coding.
5. Set thinking behavior. On K2.6, thinking is on by default and you disable it with {"type": "disabled"} in the thinking parameter. K3 accepts a top-level reasoning_effort field.
6. For tool use, keep the assistant’s reasoning_content in the conversation across steps, as the K2.6 guide recommends.

Note the K2.6 guide’s constraints: with thinking on, tool_choice must be auto or none, the built-in $web_search tool does not work in thinking mode, and sampling is fixed (temperature 1.0 with thinking, 0.6 without).

Running K2 Thinking yourself: Weights, engines and hardware

Since the hosted endpoint is gone, the open weights are the way to use K2 Thinking itself. They are published at moonshotai/Kimi-K2-Thinking on Hugging Face under the modified MIT license. The card lists vLLM, SGLang and KTransformers as supported inference engines, recommends a temperature of 1.0, and says the inference engine must support Kimi K2’s native tool-call parsing.

The checkpoints use the compressed-tensors format and can be converted to higher precision such as FP8 or BF16. At 1 trillion parameters, this is server-class hardware territory, not a laptop model; Moonshot’s pages do not give a minimum hardware spec, so check the engine’s documentation. Moonshot’s pages also do not cover Ollama, Groq or OpenRouter availability, and third-party hosts add and drop models, so check each host’s current model list. OpenRouter’s page for the model, for example, notes it is not available through EU in-region routing.

If data location matters to you, hosting open weights yourself or through a provider you choose is the main alternative to Moonshot’s own endpoint. The broader US-China policy backdrop is covered in our report on US plans to block Chinese firms from exploiting American models.

Limits and things to know before you build on it

• The Moonshot-hosted K2 Thinking is retired. Anything still pointing at it, or at K2.5, will error.
• Benchmarks are the maker’s. The scores above come from Moonshot’s model card, and the SWE-bench result trails the two rivals it lists.
• The license is modified MIT. Read the terms on the model card before commercial use; the exact modifications are stated there.
• Thinking mode has restrictions on tool choice and built-in web search, per the K2.6 guide.
• Model names change fast. Moonshot retired five model families in about a year, so pin versions and watch the changelog.

Common questions