Kimi K3 Is Here: The First Open 3T-Class Model, Now on MuiRouter

On July 17, 2026, Kimi released its new flagship, Kimi K3. The headline feature is naturally its 2.8 trillion parameters: Kimi calls it the first open model to enter the 3T class. Brute force works, and more parameters certainly help—but scale alone would not have shaken the industry quite this much.

Last night, one of the heavy hitters in my group tried Kimi K3 and immediately became an unpaid evangelist. His exact words were:

I tried it. It’s so strong. I already upgraded to the ¥199 plan—the only one with 1M. After using it, GPT feels dumb by comparison (for frontend work).

Others in the group followed his lead, started testing it, and subscribed to the ¥199 tier. By this morning, K3 was all over social media.

Kimi has put a 1M context window, native vision, long-horizon coding, tool use, and knowledge work into K3. Its API pricing is also aggressive—roughly 60% of GPT-5.6’s. Kimi is not trying to make another model that simply chats with users; it wants an agent that can keep working and finish things on its own.

We have already added kimi-k3 to MuiRouter. Keep using the same API key and /v1/chat/completions; change the model name and you can start testing. The Playground now supports image input, separates the thinking process from the final answer, and shows cached-token usage.

What exactly is Kimi K3?

According to the official Kimi K3 release and the API guide, the core specifications are:

KDA and AttnRes are the main architectural upgrades. KDA uses a more efficient attention mechanism to support long contexts, while AttnRes lets the model selectively retrieve information across network depth. Combined with a sparser MoE, Kimi says K3 delivers about 2.5× better overall scaling efficiency than K2.

Parameters and architecture matter, but developers have more immediate questions: is it good value, and can they buy it easily? GLM-5.2 coding plans are still something people have to fight to get.

The four capabilities that matter most

1. Long-horizon coding

Kimi has made long-horizon coding K3’s first battleground. Its official examples go far beyond generating a function: the model operates terminals for long periods, understands massive repositories, repeatedly runs and fixes code, builds a GPU compiler system from scratch, optimizes kernels, and even completes a 48-hour chip-design task.

These examples come from Kimi itself, so they do not automatically translate to our own projects. They do show that Kimi K3 is focused on real engineering workflows. We do not spend our days writing bubble sort; we coordinate complicated processes and implement messy business logic.

2. Vision in the loop

K3 has avoided the path taken by DeepSeek, MiMo, and GLM: it does not bolt on a separate OCR system. Kimi describes its vision as vision in the loop—the model writes code, looks at the real screenshot, then keeps adjusting based on what it sees. That is useful for frontend work, games, CAD, data visualization, and any workflow built around “make it, look at it, improve it.” Its performance last night also made the point clearly; I recommend trying it if you do frontend work.

In the MuiRouter Playground, you can upload PNG, JPEG, WebP, or GIF files and send images and text in the same message. We have not enabled video uploads yet. For quickly testing screenshot understanding, UI analysis, or image-and-text Q&A, the existing image input is already enough.

3. 1M context and Prompt caching

A 1M context window lets the model read larger repositories, more documents, or longer research material in one pass. More importantly, Kimi designed automatic caching alongside the long context: when a repeated prefix hits the cache, those input tokens are charged at a much lower rate.

This matters for coding agents. In a long session, system instructions, repository context, and message history are sent repeatedly. With a high cache-hit rate, only a small amount of new content needs to be billed at the full input rate. Kimi says its official API can exceed a 90% cache-hit rate on coding workloads, though the real number still depends on your prompt structure and session management.

4. Tool use and structured output

K3 supports custom tools, tool_choice, JSON Mode, and Structured Output. These are now essential agentic capabilities: choose a tool, generate its arguments, read the result, and continue reasoning.

One important detail is that K3 depends on complete thinking history. In complex multi-turn tool use, the agent harness should pass the previous reasoning_content back unchanged. Switching from another model to K3 in the middle of a session can also make behavior unstable. Simple chat users do not need to worry about this, but long-running agents cannot skip this compatibility detail.

Benchmarks? Read them, then move on

Kimi’s published results place K3 in the frontier-model range, and it approaches or beats more expensive closed models in parts of coding and knowledge work. But most of these numbers come from Kimi’s own evaluations. Kimi also says plainly that K3’s overall experience still trails Claude Fable 5 and GPT-5.6 Sol. It lists two limitations itself: K3 is sensitive to thinking history, and it can be overly proactive during long tasks, making decisions the user never explicitly authorized.

API pricing

Moonshot AI’s official Kimi K3 pricing is:

Cache-hit input costs one-tenth of normal input, making it the most valuable part of K3’s cost structure. At the same time, K3 currently always uses the highest thinking effort, so even simple tasks can generate substantial reasoning tokens. It is better suited to work that genuinely needs reasoning, long context, or multi-step execution than to replacing every inexpensive small model.

What MuiRouter supports

kimi-k3 on MuiRouter now supports:

• A 1M-token context window;
• PNG, JPEG, WebP, and GIF image input;
• Always-on max thinking;
• Streaming output with separate thinking and final-answer displays;
• Custom tools, tool_choice, and response_format;
• Prompt caching, with total input, cached input, and output tokens shown in the Playground;
• The OpenAI-compatible /v1/chat/completions interface.

The existing kimi-k2.6 remains available as a lower-cost independent option. K3 is a better fit for difficult, long-running, and visual tasks. If you do not need deep reasoning, there is no reason to upgrade just because the model is new.

The simplest call looks like this:

curl https://api.muirouter.com/v1/chat/completions \
  -H "Authorization: Bearer $MUI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {"role": "user", "content": "Analyze this project and propose the next implementation plan."}
    ],
    "reasoning_effort": "max",
    "max_completion_tokens": 16384,
    "stream": true
  }'

If you already use the OpenAI SDK, point the Base URL to https://api.muirouter.com/v1 and change the model to kimi-k3. Image input continues to use standard image_url content parts, so there is no new interface to learn.

Bottom line

It is exciting to see Chinese models keep improving and open models keep pushing forward. Kimi K3 is the first model to put an open 3T-class architecture, a 1M context window, native vision, and long-horizon agent capabilities into one competitive package. More providers will almost certainly offer K3 soon.

Kimi K3 probably cannot completely replace Claude Fable 5 or GPT-5.6 Sol. Kimi admits there is still an overall experience gap; max thinking adds latency and output cost; and complex agents must preserve thinking history correctly. None of that stops us from moving some work to K3—especially frontend development.

You can now select Kimi K3 in the MuiRouter Playground and test it with your own code, images, and real tasks. It is still one API key and one request format. You only need to change the model name.