What Is Kimi K3?
Kimi K3 is the newest flagship large language model from the Chinese AI lab Moonshot AI, and it has quickly become one of the most talked-about open-weight releases of 2026. The model scales to roughly 2.8 trillion total parameters, making it the largest open-weight model released to date, while still activating only a small fraction of those parameters per request through a sparse mixture-of-experts design. Kimi K3 pairs this scale with a one-million-token context window and native image understanding, so it can read very large codebases, documents, or screenshots in a single pass.
What Makes Kimi K3 Different
Two architectural choices set Kimi K3 apart from earlier Kimi releases. The first is Kimi Delta Attention, a hybrid linear-attention mechanism paired with an approach Moonshot calls Attention Residuals, both aimed at keeping information flowing cleanly through very long sequences and very deep networks. The second is an updated mixture-of-experts routing scheme, called Stable LatentMoE, which activates only 16 of 896 available experts for any given token. Moonshot reports that these changes roughly double the compute efficiency compared with the earlier Kimi K2 generation, translating into stronger reasoning and coding performance without a matching rise in serving cost.
Kimi K3 is built mainly for two kinds of work. The first is long-horizon coding, where it can operate across large repositories, run terminal tools, and use screenshots as feedback while building software. The second is knowledge work, covering research, analysis, and multi-step agent tasks that span long documents.
How to Get Started with Kimi K3
There are three common ways to try Kimi K3, depending on whether you just want to chat with it or build something on top of it.
Step 1: Chat with Kimi K3 in the Browser
The simplest starting point is Moonshot’s own consumer app at kimi.com, where Kimi K3 powers the default chat experience. No setup is required beyond creating an account, which makes it a good way to test the model’s reasoning, coding, and image-understanding abilities before committing to an API integration.
Step 2: Call Kimi K3 Through the Kimi API Platform
Developers who want to build with Kimi K3 can use Moonshot’s Kimi API Platform, which mirrors the OpenAI SDK request format. In practice this means creating an account and generating an API key, storing that key as an environment variable rather than hard-coding it, installing the OpenAI Python or Node.js SDK, and then pointing the client at Moonshot’s API base URL with the model name set to kimi-k3. From there, a chat completion request works the same way it would against any OpenAI-compatible model.
Kimi K3 runs in a thinking mode at all times, and you can control how much internal reasoning it performs through a reasoning-effort setting of low, high, or max. Streaming responses separate the model’s reasoning from its final answer, and the model also accepts base64-encoded images for vision tasks, though it does not yet support public image URLs directly.
Step 3: Access Kimi K3 Through a Routing Provider
If you would rather not manage a direct Moonshot account, aggregator platforms such as OpenRouter also expose Kimi K3 through a single API, along with usage and latency data so you can compare it against other models before committing.
What Kimi K3 Costs
Pricing through Moonshot’s own platform runs around three dollars per million input tokens and fifteen dollars per million output tokens. Prompt caching can cut the effective input cost substantially for workloads that reuse a lot of context, such as repeated questions against the same large document.
Who Should Try Kimi K3
Kimi K3 tends to make the most sense for teams building autonomous coding agents that need to hold an entire codebase in context, for long-document research and analysis, and for multimodal workflows that mix code, screenshots, and text in the same task. Smaller or latency-sensitive projects may still be better served by Moonshot’s lighter Kimi K2 family.
When Do the Full Model Weights Ship?
As of this writing, Moonshot has made Kimi K3 available through its API and chat product, but the open-weight release for self-hosting is scheduled for July 27, 2026, alongside a technical report covering the model’s architecture, training, and benchmark results.
Frequently Asked Questions
Is Kimi K3 open source?
Yes. Moonshot describes it as the first open-weight model in the 3-trillion-parameter class, with downloadable weights following shortly after the API launch.
How does Kimi K3 compare to Kimi K2?
Moonshot reports notably better efficiency and long-horizon coding performance than K2, thanks to the new attention design and sparser expert routing, though independent benchmarks are still emerging.
Do I need special hardware to use Kimi K3?
Not if you access it through the chat app, the Kimi API, or a routing provider like OpenRouter. Self-hosting the open-weight release will require substantial GPU infrastructure given the model’s size.
Resources and Further Reading
The links below go directly to Moonshot’s official documentation and platforms referenced in this guide.
- Kimi K3 Quickstart Guide – official developer docs covering setup, thinking effort, streaming, and vision input
- Kimi K3 Pricing – current input and output token rates
- Kimi Chat App – try Kimi K3 for free in your browser
- Kimi K3 on OpenRouter – third-party pricing, uptime, and benchmark comparisons
- Moonshot AI – the company behind the Kimi model family
Final Thoughts
Kimi K3 is a good example of how quickly the open-weight AI landscape is moving: a trillion-plus-parameter model with long-context and vision support that you can start testing today through a simple API call, well before its full weights are even public. If you are evaluating it for coding agents or knowledge-heavy workflows, the fastest path is to try the chat app first, then move to the API once you know which use case fits.
