Kimi K3: the AI model taking on the world’s most advanced systems

In 2023, Yang Zhilin, a researcher specializing in large language models, founded the Chinese startup Moonshot AI. The company made a strong entrance into the Chinese market thanks to the ability of its language model, Kimi, to process very long documents, a feature that was quite unusual at the time. Since then, it has evolved to the point of being able to compete directly with models such as ChatGPT, Gemini, and DeepSeek, becoming one of China’s most promising models and drawing strong interest from the rest of the world for its potential to become a benchmark in the global race to lead artificial intelligence.

And it certainly has become one.

Kimi K3, hunting down Sol and Fable

Three years after its founding, Moonshot AI has just launched Kimi K3, a new model that is generating major buzz among AI experts worldwide and has already claimed first place in the LLM Arena for web interface creation, even outperforming the newly arrived and already famous GPT 5.6 Sol and Claude Fable 5.

Let us take a look at its main features.

1.An open model competing with the best closed models

Unlike many of its direct competitors, Moonshot has built an open model, allowing researchers, companies, and developers to run and adapt it on their own infrastructure. In addition, according to the benchmarks published by the company, Kimi K3 delivers results comparable to the most advanced OpenAI and Anthropic models in programming, placing it on the front line of today’s AI.

2.A model with (almost) three trillion parameters and one million tokens of capacity per conversation

Kimi K3 has 2.8 trillion parameters, making it, according to Moonshot, the first open model to reach this scale, known as 3T. The model’s full weights will be published before July 27, 2026.

It can process up to 1 million tokens in a single conversation or task, which allows it to work with enormous codebases, lengthy documents, or multiple files without losing context.

3. A lot of power with fewer resources

Kimi K3 uses a Mixture of Experts (MoE) architecture, similar to having a team of 896 specialists where only the 16 best suited for each task step in. Thanks to this, it combines the capacity of a gigantic model with a much lower runtime cost, since it does not need to activate all of its parameters for every response.

In addition, Moonshot states that this is a much more efficient model than its predecessors: improvements in architecture, training, and data make it possible to turn computing power into useful capability with efficiency approximately 2.5 times greater than K2

4. An expert at programming

Although it is a general-purpose model, Kimi K3 stands out in programming. As the company explains, “when operating with minimal human supervision, it can sustain long engineering sessions and navigate massive repositories. Kimi K3 also excels at tasks that combine software engineering with visual reasoning: it leverages screenshots and visual elements to optimize game development, frontend work, and CAD.”

In fact, the company specifically highlights its capabilities when it comes to developing software such as video games, thanks to its 3D reasoning, programming, and computer vision capabilities. In the first few hours after its launch, we have already seen some examples of this:

Kimi K3 combines strong 3D reasoning, coding, and vision capabilities to turn concepts, images, and videos into fully playable interactive experiences.

Kimi K3 achieves true «vision in the loop» by seamlessly iterating between code and live screenshots pic.twitter.com/E9nw6yivcT

— Kimi.ai (@Kimi_Moonshot) July 16, 2026

5. Agents

Kimi K3 was built with the agentic era in mind. Moonshot emphasizes that K3 does not just answer questions, but is optimized to carry out complete work, solving complex tasks through agents capable of executing long processes and collaborating with external tools.

A threat to OpenAI and Anthropic?

The impact of Kimi K3’s launch among experts and educators in the field has been considerable, and most are talking about the remarkable ability of an open model like this one to seriously compete with the most advanced models of the moment, Sol and Fable. In any case, Kimi’s main strengths are still concentrated in the coding space, and it remains behind its competitors in other areas.

In fact, the launch announcement itself explains, “while its overall performance still falls short of the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated state-of-the-art performance across our full test suite, consistently outperforming the other models evaluated.”

This is something we can confirm by taking another look at the LLM Arena. Out of the 12 individual rankings in which users from around the world evaluate models, Kimi K3 still falls outside the top 10 in several highly relevant categories such as “agents,” “search,” “text-to-image generation,” and “image editing,” while it stands out in areas such as “web development” (1st) and “text generation” (9th).

In the video category it still does not have a rating, although Moonshot itself has highlighted that Kimi K3 edited its own teaser video from 56 original clips, “handling clip selection, motion-synced cuts, precise rhythm synchronization, audio processing, and multiple rounds of review.”

You can take a look and decide what you think 😉

Meet Kimi K3 pic.twitter.com/ou00Av3VoS

— Kimi.ai (@Kimi_Moonshot) July 16, 2026

In any case, Elon Musk does not seem too concerned about its arrival.

Whether or not it is an immediate threat to OpenAI or Anthropic, open models are arriving ready to compete head-to-head with the most advanced proprietary systems on the market, and China wants to play a leading role in that race.

Image: ChatGPT