Every Kimi AI Model Explained and Compared (Jul, 2026)

TL;DR: Kimi is Moonshot AI's model family. The flagship, Kimi K2.6, is a one-trillion-parameter open-weight model that only activates 32 billion parameters per token, so it runs cheap. It costs $0.60 input and $2.50 output per million tokens on the API, with a 256K context window, and the weights are open under a modified MIT license. There are also coding and reasoning variants.

Kimi is the open-weight model that ties the best closed models on coding at a fraction of the price. The flagship, Kimi K2.6, is a Mixture-of-Experts model with one trillion total parameters but only 32 billion active per token. That design keeps it strong and cheap at once. Around it sit Kimi K2.7 Code for the hardest coding, Kimi K2 Thinking for deep reasoning, and Kimi K2 Instruct for general work. All are open weight and also available on the Moonshot API.

The numbers in this guide come from Moonshot’s own model and pricing pages, plus the Kimi model cards on Hugging Face, read in July 2026. API prices are per one million tokens, input first, then output. We cover the models, the open weights, and how to pick. We do not cover fine-tuning recipes or the free consumer chat app, which are separate topics.

Why does this matter now. Kimi gives you near-frontier coding and agent quality as open weights, so you can self-host with full data control, or use a very cheap API. But running a one-trillion-parameter model needs real infra. This guide helps your team pick between the API and self-hosting, the same way we match engineers to roles.

The Kimi models at a glance

Here is the lineup, with the model name your engineers will call. Use this table as your reference. Everything below explains the details. Prices are input then output, per one million tokens.

The Kimi lineup in plain English

Kimi keeps the lineup focused. There is one flagship, a coding variant, a thinking variant, and a lighter instruct model, all built on the same K2 architecture. The older Moonshot V1 line is being retired. Once you see that the variants share one base model, the naming stops feeling confusing.

One trillion parameters, 32 billion active

The headline is the architecture. Kimi K2 is a Mixture-of-Experts model with one trillion total parameters, but it only activates 32 billion of them on each token. That is how it delivers frontier-class quality at a low price: you get the knowledge of a huge model but pay to run only a small slice per token. The chart below shows the gap between total and active parameters.

K2.7 Code, K2 Thinking, and K2 Instruct

The variants tune the base model for a job. K2.7 Code is Moonshot’s strongest coding model, built for agentic, multi-step coding, and it ties the best closed models on code benchmarks at a fraction of the cost. K2 Thinking is the reasoning variant, which thinks step by step and posts state-of-the-art scores on hard exams. K2 Instruct is the lighter general model for simpler work at a lower cost. All share the 256K context window.

Open weights and self-hosting

Kimi K2 is open weight under a modified MIT license, which allows commercial use and self-hosting. You can download the weights from Hugging Face and run them yourself. The models even ship with native INT4 quantization, which shrinks the memory needed to serve them, though a one-trillion-parameter model still needs a serious GPU setup.

Self-hosting gives you full data control and zero per-token cost, which matters for private data and very high volume. The trade is the infra. Serving a large MoE model efficiently, with the right quantization and batching, is a real skill. It is a job for strong AI and machine learning engineers.

Pricing and the agentic edge

On the API, Kimi K2.6 costs $0.60 input and $2.50 output per million tokens. That is cheap for a model that ties frontier coding quality, which is why cost-sensitive teams pick it for heavy coding and agent work. Cache discounts and self-hosting push the cost lower still. The snapshot below sums up the lineup.

The other edge is agents. Kimi is built for long-horizon work, where an agent runs many steps toward a goal without losing track. It also has a swarm mode that runs several agents together on one task. For teams building coding agents or autonomous workflows, that focus is a real draw. Getting the most from it takes engineers who have shipped agents before.

Context windows and thinking mode

Kimi K2 offers a 256K token context window and up to 256K max output, roughly 190,000 words each way. That is enough for large codebases, long document sets, and extended agent runs that keep a lot of history. It is smaller than the 1M some rivals offer, but large enough for almost all real work.

The K2 Thinking variant adds explicit reasoning. It thinks step by step before answering, which lifts quality on hard math, code, and logic, at the cost of more output tokens. For simple tasks, the base K2.6 or the lighter Instruct model is faster and cheaper. An engineer who picks the right variant runs Kimi far cheaper at the same quality. You can read more in our breakdown of Claude versus ChatGPT for coding.

Which model should you use

Start with Kimi K2.6 as the default. Use K2.7 Code for heavy coding and agent work. Use K2 Thinking when a task needs deep reasoning. Drop to K2 Instruct for simple, high-volume work. Self-host the open weights when you need data control or run very high volume. The table below maps common jobs to a model.

A simple rule helps. Prototype on the API, measure your real volume, then decide if self-hosting pays off. We worked with a team that ran a coding agent on a frontier US API. Moving the bulk of the coding to Kimi K2.7 Code cut their model bill by more than half with no drop in quality, because Kimi matches top models on code at a much lower price.

Older Kimi models

Moonshot ships fast, and older versions are being folded into the K2 line. If your code calls an older model, plan to move it. The table shows the main older models and what to use instead.

A move is usually small. On the API you change the model name. For open weights you download the new checkpoint and update your serving config. The Moonshot API docs and the Kimi model cards on Hugging Face show every current and older version in one place, so check them before you migrate.

What this means for hiring

Knowing the lineup is step one. Getting value from it is the harder step. With Kimi the biggest levers are picking the right variant, serving a large MoE model efficiently, and building clean agentic flows. The teams that win do this well, not the ones on the biggest model. That is engineering, not just a model choice.

Engineers who can deploy and optimize large open models are scarce and costly in the United States. A senior AI engineer there runs $12,000 to $18,000 per month all in. The same skill from a pre-vetted engineer in Vietnam or the Philippines often costs $3,000 to $6,000 per month. The work is the same. The bill is not. Many of our clients build their whole Kimi stack with an AI model deployment engineer from our network.

If you want to go deeper on the wider picture, our library of articles on AI tools and trends covers adoption, cost, and the open versus closed model debate. Read that next to see how fast this market is moving and why the right team matters more each quarter.

Build your Kimi stack with the right team

Kimi gives you frontier coding and agent quality as open weights, at a low price. The model is the easy part. The team that deploys and ships with it is what creates value. We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.