Kimi K3: Detailed Review of the New Chinese AI Model

Key Takeaways

Kimi K3 is a new AI model from the Chinese company Moonshot AI, launched on July 16, 2026. It is built on a Mixture of Experts architecture with 2.8 trillion parameters, yet only 16 out of 896 experts are activated per token, making the model more efficient. Kimi K3 supports text and image processing, handles up to 1 million tokens per request, and is available via the Kimi app, web version, Kimi Work, and API.

Architecture: What’s Inside

Kimi K3 is a Mixture-of-Experts model: instead of a single massive network, each request is processed by only 16 out of 896 specialized “expert” blocks. This architecture enhances efficiency in complex tasks, allowing a model with a nominal size of 2.8 trillion parameters to operate significantly faster and cheaper than a fully “dense” model of the same size.

The key technological feature is Kimi Delta Attention, a hybrid linear attention mechanism. It is designed for more efficient handling of long prompts — as context grows in standard models, intermediate computations increase sharply, while KDA, along with an additional technique called Attention Residuals, aims to mitigate this issue.

Model Capabilities

Text and image processing — the model understands visual content directly, without an intermediate OCR layer.

Coding and agentic tasks — according to independent benchmarks by Artificial Analysis, the model performs exceptionally well in programming tasks and working with large repositories, debugging, and iterative work with logs and tests.

Demonstration of autonomy — in Moonshot’s presentation materials, a case was shown where the model autonomously designed a functional microchip based on its own architecture over 48 hours, without human intervention at any stage of development.

Subscription in the Kimi App

Adagio — a free, unlimited basic chat with file uploads and web search, but only about 6 agentic credits.

Paid tiers, Moderato → Vivace — ranging from $19 to $199 per month — for agentic and coding tasks.

Top-up credits for API via the website — starting at $28, with 10–30% bonus credits during the launch period.

App subscription and API are two separate, unrelated accounts. Paying for a subscription does not provide API call credits, and vice versa.

Important: On July 19, 2026, Moonshot paused registration for new paid subscriptions due to demand exceeding available computing resources. Existing subscribers and API users were unaffected.

Who Is Kimi K3 For?

Consider it if you need:
— to work with very long contexts without extra charges for prompt length
— agentic and coding scenarios where accuracy and handling large repositories are critical
— multimodal tasks within a single request.